Skip to the data
THEBRRRTHE BUILDOUT
THE AI BUILDOUT / A SOURCE-BACKED FIELD GUIDE

Bigger buildout. Smarter models.

Two curves reshaping AI. The capital going in. The capability becoming affordable.

RESEARCH REVIEW · OCT 26, 2026Common spending window Jun 2026 · Meta Q3 separately reportedECI vintage Oct 26 · record Sep 22 · budget offers Jul
THE CRITICAL SIGNALS

Three numbers decide this.
Everything else is detail.

The buildout runs on one chain. These are the three places it can break — and what each one says right now.

SIGNAL 01 · THE ONE THAT MATTERS MOST

Lab revenue.

Revenue funds the next wave of compute. Watch how fast the two labs are growing.

Revenue is multiplying. Fast.

ANNUALIZED REVENUE · USD

Anthropic>740×

$87M → >$65B Jan 2024 – Jul 2026

OpenAI~25×

$2B → ~$50B End 2023 – Sep 2026

$100M$1B$10B$100B202420252026Anthropic · Start of Jan 2024 · $87MAnthropic · Dec 2024 · $1BAnthropic · Early Mar 2025 · $1.4BAnthropic · End Q1 2025 (Mar 31) · $2BAnthropic · 30 May 2025 · ~$3BAnthropic · Early Jul 2025 · $4BAnthropic · End Jul 2025 · ~$5BAnthropic · 21 Oct 2025 · $7BAnthropic · End of 2025 · >$9BAnthropic · 12 Feb 2026 · $14BAnthropic · Early Mar 2026 · >$19BAnthropic · 6 Apr 2026 · >$30BAnthropic · "Earlier in May" 2026 — date imprecise · >$47BAnthropic · End of Jul 2026 · >$65BOpenAI · Late Aug 2023 · >$1BOpenAI · Mid-Oct 2023 · $1.3BOpenAI · end 2023 · $2BOpenAI · Jun 2024 · $3.4BOpenAI · Aug 2024 · $3.6BOpenAI · Sep 2024 · ~$4BOpenAI · end 2024 · $6BOpenAI · 9 Jun 2025 · $10BOpenAI · End Jul 2025 · $12BOpenAI · Aug 2025 · $13BOpenAI · Mid-Dec 2025 · >$19BOpenAI · end 2025 · >$20BOpenAI · End Feb 2026 · >$25BOpenAI · End Jul / 13 Aug 2026 · >$40BOpenAI · Late Sep 2026 · ~$50B
$100M$1B$10B$100B202420252026Anthropic · Start of Jan 2024 · $87MAnthropic · Dec 2024 · $1BAnthropic · Early Mar 2025 · $1.4BAnthropic · End Q1 2025 (Mar 31) · $2BAnthropic · 30 May 2025 · ~$3BAnthropic · Early Jul 2025 · $4BAnthropic · End Jul 2025 · ~$5BAnthropic · 21 Oct 2025 · $7BAnthropic · End of 2025 · >$9BAnthropic · 12 Feb 2026 · $14BAnthropic · Early Mar 2026 · >$19BAnthropic · 6 Apr 2026 · >$30BAnthropic · "Earlier in May" 2026 — date imprecise · >$47BAnthropic · End of Jul 2026 · >$65BOpenAI · Late Aug 2023 · >$1BOpenAI · Mid-Oct 2023 · $1.3BOpenAI · end 2023 · $2BOpenAI · Jun 2024 · $3.4BOpenAI · Aug 2024 · $3.6BOpenAI · Sep 2024 · ~$4BOpenAI · end 2024 · $6BOpenAI · 9 Jun 2025 · $10BOpenAI · End Jul 2025 · $12BOpenAI · Aug 2025 · $13BOpenAI · Mid-Dec 2025 · >$19BOpenAI · end 2025 · >$20BOpenAI · End Feb 2026 · >$25BOpenAI · End Jul / 13 Aug 2026 · >$40BOpenAI · Late Sep 2026 · ~$50B
AnthropicOpenAILog scale · each gridline = 10× the revenue

Tap a marker for its source. Arrows mark reported minimums; horizontal marks show date ranges.

Revenue history + sources · 29 observations

Reported annualized run rates, not full-year revenue. Company definitions differ. Lines connect observations, not measured values between them. Reviewed 2026-10-16; latest observations: Anthropic July 2026; OpenAI late September 2026. Dates and accounting bases differ.

Anthropic starts at company-stated $87M; conflicting October 2023 $100M excluded. OpenAI year ends use its company-published $2B/$6B/$20B+ ladder instead of conflicting or ambiguous press estimates. August 2024 $3.6B is calculated from reported $300M monthly revenue ×12. Date ranges preserve month-only and earlier-this-month observations; no interpolated observations. Forecasts and segment revenue excluded.

Anthropic
OpenAI
ILLUSTRATIVE LAB SCALE · DIFFERENT DATES / BASES~$115Bdifferent dates: Anthropic Jul 31; OpenAI late Sep 2026 · reported by Bloomberg and Reuters / FT, October 2026
HYPERSCALER CAPEX BEING SPENT TO SERVE IT$510.7Btrailing twelve months ended June 30, 2026 · four companies, as filed

About 4.4× the illustrative annualized lab scale—not actual annual sales, AI-only capex, or matched periods.

AS OFdifferent dates: Anthropic Jul 31; OpenAI late Sep 2026
NEXT UPDATENo scheduled release

Run rates different dates: Anthropic Jul 31; OpenAI late Sep 2026, reported by Bloomberg and Reuters / FT. OpenAI is approximate and Anthropic a reported floor. Different dates and definitions prevent a precise combined total or ranking.

SIGNAL 02

Hyperscaler capex.

$510.7BFour companies, trailing twelve months, as filed.
2023-Q12026-Q2 · $165.1B

Quarterly spending is 4.9× what it was 3 years ago, and the newest quarter is the largest yet.

AS OFtrailing twelve months ended June 30, 2026
NEXT UPDATEQ3 2026 results, late October 2026
SIGNAL 03 · THE EDGE CASE

Where the tokens route.

21%of tokens routed on OpenRouter now go to OpenAI, Anthropic and Google — down from 45% in Oct 2025.
Oct 2025Oct 2026 · 21%

The labs lost 24 points of routed tokens in 52 weeks — and their revenue kept multiplying anyway (Signal 01). Routing share and reported lab revenue measure different markets and periods. The combination is consistent with proliferation, but does not prove where paying work moved.

Watch for: revenue following the tokens out. Token share alone falling is what mass adoption of cheap models looks like.

AS OF · LAST COMPLETE WEEKweek of Oct 19, 20267 days old · pulled Oct 26, 2026
NEXT UPDATEWeekly cadence, re-pulled by handNot a feed. The build fails if the newest week is over 60 days old.
WHY THOSE THREE
  1. 01The two labs are the biggest marginal buyers of compute.
  2. 02They only buy more compute if their revenue grows.
  3. 03Their revenue only grows if AI usage keeps surging.
  4. 04So lab revenue is the tell. If it stalls, the buildout stalls.

The one thing that would break the chain: AI usage keeps surging but the revenue goes to cheaper models instead of the labs. Token volume shifting alone is not it. That is Signal 03, and it is why it is on this list.

Everything below is supporting detail. Capability per token, adoption indices, funding mechanics and the company-by-company splits are all genuinely interesting — and none of them decides whether the buildout continues. Read on if you want the workings.

Data + sources for the critical signalsEvery figure, dated ↗

What a run rate is. Annualized revenue run rate — a point-in-time monthly figure multiplied out to a year. Not full-year revenue, not a projection, and not audited.

Retrieval. Lab run rates transcribed from the dated press reporting named in each row, compiled 2026-09-21. Hyperscaler capex carried from funding.json (SEC EDGAR XBRL, retrieved 2026-09-21). Open-model observations transcribed from Mozilla's State of Open Source AI v1.1 and the Linux Foundation working-paper summary. Every value is committed to this file and read at build time. Nothing in this section fetches at request time.

Lab revenue. Anthropic — surpassed $65 billion at the end of July (Bloomberg, 2026-08-17; anonymous-sourced). Prior point $9B at end of 2025 (TechCrunch; anonymous-sourced). OpenAI — approximately $50B annualized run rate on company/net partner accounting in late September (Reuters / FT, 2026-10-08; press-reported approximation; attributed relay corroborated by AFP). Prior point more than $20B at end of 2025 (CNBC; company-stated forecast).

Definitions. OpenAI and Anthropic do not define revenue the same way. October reporting puts OpenAI near $50B on its own accounting, net of cloud-partner revenue shares; larger gross-partner-normalized figures are not comparable. Anthropic’s >$65B is end-July, not September. The displayed sum combines different dates and accounting definitions; it is an illustrative scale comparison, not aggregate achieved annual sales. An OpenAI spokesperson told CNBC the company had hit $10 billion annual recurring revenue, explicitly excluding licensing revenue from the Microsoft revenue share and large one-time deals. This is the only public statement of what either lab's revenue figure includes.

Hyperscaler capex. Cash purchases and additions of property and equipment for Amazon, Alphabet, Microsoft and Meta, trailing twelve months ended June 30, 2026. This value is carried unchanged from the Investor Lens dataset below and is not recomputed here. It is a company-wide financial proxy, not AI-only spending. The plotted series is the 14 quarters committed in the quarterly dataset below, 2023-Q1 through 2026-Q2; its last four quarters sum to the headline. All four report calendar third-quarter results in the second half of October. Dates are the companies' customary reporting window and are not yet confirmed filings.

Where the tokens route. Tokens routed to OpenAI, Anthropic and Google as a percentage of every token OpenRouter routed that week. Author buckets exactly as OpenRouter publishes them; every bucket the panel returns, including its residual 'others', sits in the denominator. No model-licence classification is applied — an author's tokens are counted to that author whatever the licence. Routed developer traffic only. First-party usage inside ChatGPT, Claude and Gemini never reaches this platform and is not in these numbers. Series pulled 2026-10-26 from OpenRouter; the week beginning 2026-10-26 was still incomplete and is excluded. Cadence is weekly, re-pulled by hand; the newest complete week plotted here begins Oct 19, 2026, 7 days before this dataset’s verified date of 2026-10-26. The build refuses to publish the section if that gap ever exceeds 60 days. Roughly two months. Long enough that an ordinary gap between manual refreshes does not break the build, short enough that a share this card describes as moving cannot be more than a quarter out of date on the page. Measured from the newest complete week in the series to this file's verifiedAt date. Manual October 26, 2026 pull; latest completed week October 19; incomplete October 26 bucket excluded. This is not an automatic feed. See README for the recorded pull procedure.

Open-weight share, separately. Eight of the top ten models by token volume are open weights. (Mozilla Foundation, August 2026.) Closed models account for roughly 80% of model usage and 96% of revenue. (Linux Foundation, published December 2025.) Those two readings come from different observation windows, are reported separately, and are not combined with each other or with the weekly routing series above. OpenRouter is one developer routing marketplace. The Linux Foundation authors state it captures nearly 1% of global LLM API expenditure, and it does not observe first-party ChatGPT or Claude consumer and enterprise usage at all — which is where most lab revenue is earned. Read this as a leading indicator of developer routing behaviour, not as the labs' market share.

  • Neither lab files with the SEC. There is no audited revenue figure for either company anywhere in this section.
  • Headline lab run rates have different observation dates and accounting bases: OpenAI approximately $50B late September, net partner accounting; Anthropic above $65B end July. Neither is a new October 26 measurement.
  • Run rates, full-year actuals and forward projections are three different series. Nothing on this page mixes them, and no projection is published here.
  • OpenAI approximately $50B is a reported estimate; Anthropic above $65B is a floor. These are not synchronized or audited revenue totals.
  • OpenAI's end-2025 figure was spoken in November 2025 as an expectation for year end, and is treated as achieved only by later coverage.
  • Hyperscaler cash PP&E is a company-wide proxy that includes non-AI assets. It is not a count of chips, energized data centers or AI-only spending, and it is not the compute the labs buy.
  • Hyperscaler capex and lab revenue are not two sides of one transaction. Comparing them sizes the two halves of the engine; it does not trace a payment.
  • Open-model share is observed on one developer routing marketplace that its own researchers put at nearly 1% of global LLM API spend. Usage share and revenue share come from different observation windows and are reported separately.
  • Every figure here is a committed dated observation read at build time. None of it updates on its own.
  • Both Bloomberg articles are paywalled. Their full text was not retrieved; the Anthropic figure is corroborated by the TechCrunch relay recorded alongside it.
BLOOMBERG OPENAI RUNRATE BLOOMBERG ANTHROPIC RUNRATE TECHCRUNCH ANTHROPIC 65B CNBC OPENAI END2025 CNBC OPENAI ARR DEFINITION MSFT FY2026 10K GOOGL 2026Q2 10Q AMZN 2026Q2 10Q META 2026Q2 10Q MOZILLA SOOSAI V11 OPENROUTER RANKINGS LINUXFOUNDATION OPEN MODELS OPENROUTER MARKET SHARE REUTERS OPENAI 50B
01 / THE HYPERSCALER SPENDING RAMP

Nearly 6× in six years.

5.8×

Five hyperscalers’ cash capex, 2025 vs. 2019

Annual cash spending at five large hyperscalers
  • Alphabet
  • Amazon
  • Meta
  • Microsoft
  • Oracle
EstimateHouse E0200400600800$1.0 trillion’19’20’21’22’23’24’25’26E’27E
Est.House0200400600800$1.0 trillion’19’21’23’25’26E’27E

Company-wide cash PP&E, not AI-only · 2026 is a frozen June-vintage estimate · Meta Q3 actual separately disclosed below · Oracle years run Dec–Nov

2027E = TheBRRR house estimates, not management guidance or consensus (Oct 26, 2026). Each frozen 2026 cash PP&E baseline × company-specific growth: Alphabet +30%, Amazon +25%, Meta +35%, Microsoft +25%, Oracle +20%; round each to nearest $1B, then sum. No lease or fiscal-year conversion is assumed. Oracle’s 2027 window is Dec 2026–Nov 2027; others Jan–Dec. Muted, dashed bars are house estimates; actual outcomes may differ materially. Model ledger + sources.

Source: SEC filings (10-K, 10-Q); company capex guidance for 2026

Data + sources2019–2027E · $ billions
YearAlphabetAmazonMetaMicrosoftOracleTotal
201923.516.915.113.51.670.6
202022.340.115.117.61.897.0
202124.661.118.623.23.1130.6
202231.563.631.224.86.7157.8
202332.352.727.035.26.9154.2
202452.583.037.355.610.7239.1
202591.4131.869.783.135.5411.5
2026E200.0206.8109.3138.385.0739.4
2027E260.0259.0148.0173.0102.0942.0
  • Cash purchases of property and equipment from each cash-flow statement. Finance leases excluded; Amazon before sales and incentives; Meta net of asset sales. Whole companies, not AI-only.
  • Microsoft (June fiscal year) is rebuilt from four calendar quarters. Oracle (May fiscal year) uses its four quarters ending Feb, May, Aug and Nov, so its year runs December–November.
  • 2026 estimate: Alphabet, Amazon, Meta and Microsoft are reported January–June plus this page’s outlook reference path for July–December (Alphabet’s guidance midpoint; the others hold their June quarter; Meta has since reported Q3 cash PP&E of $33.942B, separately disclosed in Outlook). Oracle is reported December–August plus September–November from its $90–95B FY2027 capex guidance, midpoint less reported spending, split equally over the remaining quarters.
  • All five 2027 values are TheBRRR house estimates, not management guidance or consensus. Reproduce: frozen 2026 chart-model cash PP&E × (1 + company-specific discretionary growth), round each company to nearest $1B, then sum. Baselines retain June-vintage assumptions (Oracle through August); these are illustrative full-year scenarios, not latest-pace forecasts. Published aggregate or lease-inclusive/fiscal-year forecasts are not a verified company-by-company cash PP&E bridge. Growth rates are editorial assumptions informed by linked investment context, not extracted source forecasts. No calendar conversion of fiscal guidance or lease bridge is invented. Actual outcomes may differ materially.
    • Alphabet: A discretionary 30% ramp reflects Alphabet’s stated significant 2027 capex increase, not a disclosed dollar target. Alphabet investment context · 2026-07-22
    • Amazon: A discretionary 25% ramp reflects continuing AWS infrastructure expansion; the approximately $220B 2026 net cash-capex headline is not substituted for gross cash PP&E. Amazon investment context · 2026-07-30
    • Meta: A discretionary 35% ramp reflects the infrastructure buildout and the reported Q3 cash PP&E acceleration ($33.942B versus $30.116B in Q2); no lease-inclusive guidance is converted to cash PP&E. Meta investment context · 2026-10-21
    • Microsoft: A discretionary 25% ramp reflects the stated FY2027 investment growth; lease-inclusive $175B CY2026 capex is not treated as cash PP&E or as CY2027 guidance. Microsoft investment context · 2026-07-29
    • Oracle: A discretionary 20% ramp reflects OCI capacity expansion; FY2027 $90–95B guidance ends May 2027, not November, and is not presented as a full 2027 calendar-window forecast. Q1 FY27 call transcript Q1 FY27 release (8-K)
  • Five companies. Signal 02 and the chapters below count four (no Oracle), so their totals differ.
  • Every value cites its SEC filing, XBRL tag and period. Full source map in the data download.
02A / INTELLIGENCE OVER TIME

Models keep getting smarter.

+39.8frontier ECI points since May 2024
Best measured intelligence · any priceBest sampled intelligence at ≤$1/M tokens

Tap a point to see the model, date and sources.

Intelligence improves over timeINTELLIGENCE · ECI POINTS ↑110120130140150160170180May 2024Jan 2025Jul 2025Jan 2026Sep 2026TIME → RELEASE / OFFER DATE
Intelligence improves over timeINTELLIGENCE · ECI POINTS ↑110120130140150160170180May 2024Jan 2025Jan 2026Sep 2026TIME → RELEASE / OFFER DATE
FRONTIER NOW168.8 ECI

Claude Opus 5.5

THE $1/M TIER NOW156.5 ECI

GPT-5.6 Luna · $0.45/M

Steps mark newly attained records, not smooth progress. ECI is an index; axis starts at 110. One benchmark vintage: 2026-10-26.

02B / WHAT ONE FIXED BUG COSTS

Fixing a real bug
used to cost real money.

151×cheaper per bug actually fixed, 4 months apart

A task here is one real bug report filed against a real open-source project. The model only scores if its patch makes that project’s own test suite pass. So this is the price of getting one genuine software fix done — not a price per word, per token, or per query.

44 runs where the test actually reported its own bill31 older runs, rebuilt from token counts × the prices of that month

Tap or click any point for the model, date, cost, share of bugs fixed and sources.

Cost per fixed bug, Oct 2024 to Mar 2026COST PER BUG ACTUALLY FIXED · US$ (LOG SCALE)$0.03$0.1$0.3$1$3$10Oct 2024Jan 2025Jul 2025Mar 2026$7.08$0.047TIME → DATE THE RUN WAS PRICED
Cost per fixed bug, Oct 2024 to Mar 2026$ PER BUG FIXED (LOG)$0.03$0.1$0.3$1$3$10Oct 2024Jan 2025Jul 2025Mar 2026$7.08$0.047TIME → DATE THE RUN WAS PRICED
Jul 2025 · DEAREST ON THE CHART$7.08

GPT 4o (2024-11-20) — fixed 21.6% of the bugs

Dec 2025 · CHEAPEST THAT DID MORE$0.047

DeepSeek V3.2 Reasoner — fixed 60.0% of the bugs

GPT 4o (2024-11-20) needed $7.08 of API spend per bug it actually fixed in Jul 2025, and fixed 21.6% of them. 4 months later DeepSeek V3.2 Reasoner fixed 60.0% of them for $0.047 each. That is 151× less money for 2.8× the work.

Each mark is one dated run of one model. Points are never joined into a trend line, because the two series are priced on different bases and the harness changed underneath them. Newest priced run on the chart: Mar 2026.

Data + sources

Intelligence timeline. Epoch AI capability scores, vintage 2026-10-26; provider price receipts and dated offers from the existing catalog. Release dates set the broad intelligence frontier; API offer dates set the budget record.

Capability source ↗

Cost per fixed bug. 44 measured points come from the official SWE-bench Verified “Bash Only” leaderboard, Jul 20, 2025 to Feb 26, 2026: each run reports its own average API spend per task. 31 reconstructed points come from Gundlach, Lynch, Mertens and Thompson, The Price of Progress, which multiplies Epoch AI’s published token counts for the benchmark by the token prices archived for that month. The reconstructed series is what reaches back into 2024, before the leaderboard published dollar costs. Every point on the chart carries its own arithmetic and source links; tap it. We divide by 484 tasks for reconstructed points — Epoch AI runs 484 of the 500 SWE-bench Verified problems, excluding 16 that do not run reliably in its infrastructure.

  • The leaderboard's dollar column is an average per task attempted, not per bug actually fixed; dividing by the resolve rate to get a per-fixed-bug figure is our step, not theirs.
  • The date on each point is when the run was submitted or the price snapshot taken. It is not the model's release date, which is carried separately on the measured points.
  • Harness versions and reasoning settings differ between rows — the measured series spans mini-SWE-agent 0.0.0 to 2.x, and the benchmark itself changed scaffold in February 2026 — so rows are not perfectly apples-to-apples.
  • Both series count API inference spend only. Neither is the all-in cost of delivering an accepted change: no human review, no failed attempts outside the run, no engineering time.
  • This is software engineering on Python repositories with passing test suites. It does not generalise to knowledge work at large.
  • October 16 source audit: newest September 1 mini-SWE-agent Gemini 3.5 Flash submission reports 71.8% resolved but publishes no API cost; cost unavailable, not plotted.
  • Reconstructed points reproduce the upstream simplified token-price model, not invoices. Cached-input discounts and separately billed reasoning tokens are not fully reconciled; first-of-month labels are price-month conventions, not exact evaluation or release dates. Do not interpret reconstructed points as measured task bills.
  • October26 limited review: current Verified payload exposes latest costed entry Feb26 Gemini3 Pro. Previously cited Sep1 Gemini3.5 Flash missing-cost row not exposed in current payload; no invented cost or new plotted point. Historical reconstruction not reaudited row-by-row.
SWE-bench leaderboard ↗Leaderboard run records ↗The Price of Progress (arXiv) ↗Reconstruction dataset ↗Epoch AI token counts ↗What SWE-bench Verified is ↗

Excluded, and why. 8 candidate runs are not plotted: 20250726_mini-v1.0.0_gemini-2.0-flash (dollar cost recorded as 0.0, which is not a usable figure); 20251021_SalesforceAIResearch_SAGE_bash_only (leaderboard entry records no dollar cost (instance_cost=None, resolved=None)); 20251209_mini-v1.17.2_devstral-2512 (leaderboard entry records no dollar cost (instance_cost=None, resolved=53.8)); 20251209_mini-v1.17.2_devstral-small-2512 (leaderboard entry records no dollar cost (instance_cost=None, resolved=56.4)); 20260901_mini-v2.4.2_gemini-3-5-flash (leaderboard entry records no dollar cost (instance_cost=None, resolved=71.8)); claude-opus-4-6_64K 02/01/2026 (dataset records no SWE-bench resolve rate, so cost per resolved task cannot be formed; dataset note: excluded their iis werid behavior here?, this has radically fewer token then the same model); claude-opus-4-5-20251101_32K 12/01/2025 (dataset records no SWE-bench resolve rate, so cost per resolved task cannot be formed; dataset note: assumed higher context version for match); 20260901_mini-v2.4.2_gemini-3-5-flash (Source metadata reports resolved fraction but no cost or instance_cost; current task cost unavailable. Source: https://github.com/SWE-bench/experiments/blob/main/evaluation/verified/20260901_mini-v2.4.2_gemini-3-5-flash/metadata.yaml).

Every point behind the cost-per-fixed-bug chart
ModelDateBasisBugs fixed$ / bug fixed
Claude 3.5 Sonnet (Oct)Oct 1, 2024Reconstructed40.6%$0.45
Claude 3.5 Sonnet (June)Nov 1, 2024Reconstructed32.0%$0.60
GPT-4o (Nov '24)Nov 1, 2024Reconstructed25.4%$3.12
o3-mini (medium) (Option2) (SWE)Jan 1, 2025Reconstructed37.8%$0.81
Gemini 2.0 FlashFeb 1, 2025Reconstructed22.0%$0.23
Claude 3.7 SonnetApr 1, 2025Reconstructed52.2%$0.72
GPT-4.1Apr 1, 2025Reconstructed41.0%$1.58
GPT-4.1 miniApr 1, 2025Reconstructed32.8%$0.31
Claude 4 OpusMay 1, 2025Reconstructed62.2%$4.20
Claude 4 SonnetMay 1, 2025Reconstructed60.6%$0.94
DeepSeek R1 0528 (May '25)Jun 1, 2025Reconstructed33.3%$0.22
Kimi K2Jul 1, 2025Reconstructed16.4%$2.25
Claude 3.7 Sonnet (20250219)Jul 20, 2025Measured52.8%$0.67
GPT 4.1 mini (2025-04-14)Jul 20, 2025Measured23.9%$1.82
GPT 4o (2024-11-20)Jul 20, 2025Measured21.6%$7.08
Llama 4 Maverick InstructJul 20, 2025Measured21.0%$1.49
Llama 4 Scout InstructJul 20, 2025Measured9.1%$1.28
Claude 4 Sonnet (20250514)Jul 26, 2025Measured64.9%$0.57
GPT 4.1 (2025-04-14)Jul 26, 2025Measured39.6%$0.37
Gemini 2.5 Flash (2025-04-17)Jul 26, 2025Measured28.7%$0.47
Gemini 2.5 Pro (2025-05-06)Jul 26, 2025Measured53.6%$0.54
o3 (2025-04-16)Jul 26, 2025Measured58.4%$0.57
o4-mini (2025-04-16)Jul 26, 2025Measured45.0%$0.47
Claude 4.1 OpusAug 1, 2025Reconstructed63.2%$4.63
GPT-5 (medium)Aug 1, 2025Reconstructed58.8%$1.21
GPT-5 mini (high)Aug 1, 2025Reconstructed59.2%$0.32
GPT-5 mini (medium)Aug 1, 2025Reconstructed57.8%$0.12
Claude 4 Opus (20250514)Aug 2, 2025Measured67.6%$1.67
Qwen3-Coder 480B/A35B InstructAug 2, 2025Measured55.4%$0.45
Qwen2.5-Coder 32B InstructAug 3, 2025Measured9.0%$0.76
GPT 5 (2025-08-07) (medium)Aug 7, 2025Measured65.0%$0.43
GPT 5 mini (2025-08-07) (medium)Aug 7, 2025Measured59.8%$0.059
GPT 5 nano (2025-08-07) (medium)Aug 7, 2025Measured34.8%$0.11
Kimi K2 InstructAug 7, 2025Measured43.8%$1.21
gpt-oss-120bAug 7, 2025Measured26.0%$0.22
GLM 4.5 (2025-08-22)Aug 22, 2025Measured54.2%$0.55
DeepSeek R1 0528 (May '25)Sep 1, 2025Reconstructed33.3%$0.15
DeepSeek V3.1 (Non-reasoning)Sep 1, 2025Reconstructed52.1%$0.35
Kimi K2Sep 1, 2025Reconstructed16.4%$1.97
Claude 4.5 Sonnet (20250929)Sep 29, 2025Measured70.6%$0.79
Claude 4.5 Haiku (Non-reasoning)Oct 1, 2025Reconstructed60.6%$0.35
Claude 4.5 Sonnet (Non-reasoning)Oct 1, 2025Reconstructed64.8%$1.08
gemini-3-pro-previewNov 1, 2025Reconstructed72.9%$3.84
gpt-5.1-2025-11-13_highNov 1, 2025Reconstructed67.8%$3.97
Gemini 3 Pro Preview (2025-11-18)Nov 18, 2025Measured74.2%$0.62
GPT 5.1 (2025-11-13) (medium)Nov 20, 2025Measured66.0%$0.46
Claude 4.5 Opus (20251101) (medium)Nov 24, 2025Measured74.4%$0.97
GPT 5.1 Codex (medium)Nov 24, 2025Measured66.0%$0.89
MiniMax M2Nov 24, 2025Measured61.0%$0.70
DeepSeek V3.2 ReasonerDec 1, 2025Measured60.0%$0.047
GLM 4.6 (T=1)Dec 1, 2025Measured55.4%$0.17
gpt-5.2-2025-12-11_highDec 1, 2025Reconstructed73.8%$0.27
Kimi K2 ThinkingDec 10, 2025Measured63.4%$0.69
GPT 5.2 (2025-12-11)Dec 11, 2025Measured69.0%$0.39
GPT 5.2 (2025-12-11) (high)Dec 11, 2025Measured71.8%$0.72
fireworks/kimi-k2p5Jan 1, 2026Reconstructed73.7%$0.93
claude-opus-4-6_32KFeb 1, 2026Reconstructed75.6%$0.44
claude-sonnet-4-6_32KFeb 1, 2026Reconstructed75.2%$0.55
fireworks/kimi-k2p5Feb 1, 2026Reconstructed73.7%$0.78
gemini-3.1-pro-previewFeb 1, 2026Reconstructed75.6%$6.35
glm-5Feb 1, 2026Reconstructed72.1%$2.53
Claude 4.5 Haiku (high)Feb 17, 2026Measured66.6%$0.50
Claude 4.5 Opus (high)Feb 17, 2026Measured76.8%$0.98
Claude 4.5 Sonnet (high)Feb 17, 2026Measured71.4%$0.92
Claude 4.6 OpusFeb 17, 2026Measured75.6%$0.73
DeepSeek V3.2 (high)Feb 17, 2026Measured70.0%$0.64
GLM 5 (high)Feb 17, 2026Measured72.8%$0.73
GPT 5 miniFeb 17, 2026Measured56.2%$0.084
GPT 5.2 (high)Feb 17, 2026Measured72.8%$0.65
Gemini 3 Flash (high)Feb 17, 2026Measured75.8%$0.47
Kimi K2.5 (high)Feb 17, 2026Measured70.8%$0.21
MiniMax M2.5 (high)Feb 17, 2026Measured75.8%$0.097
GPT 5.2 CodexFeb 19, 2026Measured72.8%$0.62
Gemini 3 ProFeb 26, 2026Measured69.6%$1.38
fireworks/kimi-k2p5Mar 1, 2026Reconstructed73.7%$0.70
Compare intelligence against priceOptional scatter view
02 / THE INTELLIGENCE FRONTIER

More intelligence.
A wider price range.

7.6×

CHEAPER FOR GPT-5’S INDEX SCORE
$3.44 → $0.45 / MILLION TOKENS

PRICIEST OFFER · APR 24, 2026GPT-5.5 Pro$67.50 / M
SAME SCORE, 7.6× CHEAPERPRICES FROM $0.13 TO $67.50/MGPT-5’s score: $3.44/M → $0.45/M in 12 months.24 dated offers over 28 months · 8 at or under $1/M, 16 above110120130140150160170180$0.10$0.30$1$3$10$30$100$1/M CEILING▲ SMARTER · ECI POINTSPRICE · $ PER MILLION TOKENS (LOG)◀ CHEAPER3 frontier releases163.4–168.8 ECI · $20/MNEW PRICE HIGHS ▶◀ 7.6× CHEAPERGemini 1.5 Flash · May ’24122.7 ECI · $0.53/MGemini 1.5 Flash · Sep ’24129.3 ECI · $0.13/MGPT-5150.0 ECI · $3.44/MGPT-5.5 Pro162.9 ECI · $67.50/MGPT-5.6 Luna156.5 ECI · $0.45/MGPT-6 Astra168.8 ECI · $20/MIndex axis starts at 110, not zero · price axis is logarithmicMay 13, 2024Sep 3, 2026Colour + size = offer date
SAME SCORE, 7.6× CHEAPERPRICES FROM $0.13 TO $67.50/MGPT-5’s score: $3.44/M → $0.45/M in 12 months.24 offers · colour + size show date110130150170180$0.10$0.30$1$3$10$30$100$1/M CEILING▲ SMARTER · ECI POINTSPRICE · $ / M TOKENS (LOG)◀ CHEAPERNEW PRICE HIGHS ▶Gemini 1.5 Flash · Sep ’24129.3 ECI · $0.13/MGPT-5.6 Luna156.5 ECI · $0.45/MGPT-6 Astra168.8 ECI · $20/MIndex axis starts at 110, not zero · price axis is logarithmicMay 13, 2024Sep 3, 2026Colour + size = date
Offer at ≤$1/MOffer above $1/MNew price highsTap or focus a point for its model
Capability and price for every plotted model offer, oldest first. GPT-5’s index level cost $3.44 per million tokens in Aug 7, 2025 and $0.45 by Jul 30, 2026: 7.6× cheaper for the same score. Over the same board the most expensive offer rose from $7.50 to $67.50 per million tokens, 9.0× more, with no new high since Mar 5, 2026. Launch prices for expanded flagship coverage; dated budget offers include price cuts. Current-only quotes are excluded.
ModelProviderOffer dateECI90% intervalBlended USD per million tokensPrice tier
GPT-4o (May 2024)OpenAIMay 13, 2024128.99124.62–131.237.5above $1/M
Gemini 1.5 Flash (May 2024)Google DeepMindMay 30, 2024122.68116.30–124.770.525at or under $1/M
Claude 3.5 SonnetAnthropicJun 20, 2024130.00not published6above $1/M
GPT-4o miniOpenAIJul 18, 2024126.46120.11–128.490.2625at or under $1/M
o1-miniOpenAISep 12, 2024135.70132.36–137.175.25above $1/M
Gemini 1.5 Flash (Sep 2024)Google DeepMindSep 24, 2024129.33123.75–131.200.1313at or under $1/M
o1OpenAIDec 17, 2024141.72139.71–142.9426.25above $1/M
DeepSeek-R1DeepSeekJan 20, 2025138.82136.04–140.210.96at or under $1/M
o3OpenAIApr 16, 2025146.73144.83–148.5217.5above $1/M
Gemini 2.5 Pro (Mar 2025)Google DeepMindApr 30, 2025143.96141.74–146.843.4375above $1/M
Gemini 2.5 Flash (May 2025)Google DeepMindMay 20, 2025141.43139.18–142.820.9875at or under $1/M
o3-proOpenAIJun 10, 2025147.33145.56–150.1035above $1/M
GPT-5OpenAIAug 7, 2025150.00not published3.4375above $1/M
GPT-5 miniOpenAIAug 7, 2025145.34143.33–147.110.6875at or under $1/M
GPT-5 ProOpenAIOct 7, 2025150.59148.94–154.1841.25above $1/M
Gemini 3 ProGoogle DeepMindNov 18, 2025153.07151.06–155.424.5above $1/M
GPT-5.2 ProOpenAIDec 11, 2025156.33153.89–159.9857.75above $1/M
MiniMax-M2.5MiniMaxFeb 12, 2026146.59142.21–147.900.825at or under $1/M
GPT-5.4 ProOpenAIMar 5, 2026159.57157.10–162.7067.5above $1/M
GPT-5.5 ProOpenAIApr 24, 2026162.91160.04–166.6067.5above $1/M
Claude Fable 5AnthropicJun 9, 2026163.42160.53–168.1820above $1/M
GPT-5.6 LunaOpenAIJul 30, 2026156.50154.19–159.260.45at or under $1/M
Claude Fable 5.1AnthropicSep 1, 2026166.49163.47–171.3520above $1/M
GPT-6 AstraOpenAISep 3, 2026168.81165.45–174.5120above $1/M
GPT-5 (reference)OpenAIAug 7, 2025150.00not published3.4375above $1/M
THEBRRR READ

Yesterday’s frontier, a fraction of the price.

7.6×cheaper for GPT-5’s index score
9.0×rise in the highest sampled list price
514×spread, $0.13 to $67.50 / M

Matching GPT-5’s 150 ECI cost $3.44/M in Aug 7, 2025 and $0.45/M by Jul 30, 2026—7.6× cheaper for the same score. Over the same board the most expensive offer went $7.50 → $67.50/M, 9.0× dearer, and no offer has priced above $67.50/M since Mar 5, 2026. GPT-6 Astra now tops the board at 168.81 ECI for $20/M.

Strong capability is becoming affordable. The highest price is not the highest intelligence score.

Epoch AI · Oct 26, 2026. One index vintage; overlapping 90% intervals. Points, not “% smarter.” Dated list prices, not task costs.

SHOW CONTEXT
Explore the model milestones & pricesData + sources ↗
Jul 30, 2026 · VERIFIED OFFERGPT-5.6 Luna156.50 ECI · $0.45 / M blended tokens
Input $0.20 / MOutput $1.20 / M75% input / 25% output · uncached, standard rates
Price announcementHow to read this

A historical record from 24 selected dated offers, not a complete market survey or continuous availability. ECI is a benchmark index—not “times smarter.” Token prices are not task costs.

The charts show the shift. The research follows the opportunity. Who captures the spending, where cheaper AI unlocks demand, and what changes next.

Understand the implications
What the capability–price scatter measuresMethod + benchmark and price sources

A price per million tokens, not a task budget. Each point is one dated API offer at a 75% uncached text input / 25% billed output mix, including reasoning tokens. The $1/M gridline is a reading aid that splits the 8 sub-ceiling offers from the 16 above it; it is not a selection filter. Different models use different numbers of tokens to solve a task, so a lower per-token price is not automatically a lower cost per job. No batch, cached, flex, priority, subscription, promotional, enterprise or regional-residency rates. Every reasoning model here bills reasoning/thinking tokens as standard output at the same rate, stated explicitly by both providers. No model in this set prices reasoning separately. Cheapest published tier taken for each model. That tier runs to 128K on older OpenAI models, 200K on Gemini 2.5 Pro and Gemini 3 Pro, and 272K on GPT-5.4 and later. Every quoted rate satisfies the under-128K short-context basis, but the headroom bought differs by provider.

One capability ruler. Epoch Capabilities Index combines benchmark results. All scores use the same Oct 26, 2026 vintage, retrospectively applied to historical models. Points and 90% intervals are index estimates, not percentages or ratios of intelligence. The vertical scale is 110–180, with no meaningful zero.

How to read the scatter. Each mark is one dated offer at its own blended price, so the horizontal axis is a real price axis: circles are the 8 offers at or under $1 per million tokens, diamonds the 16 offers above it. Colour and size run with the offer date, May 13, 2024 to Sep 3, 2026. Price uses a logarithmic axis because the range spans $0.10 to $100; a linear axis would stack every cheap model on the margin. Epoch’s broader 270-model release record carries no price column, so it is not plotted here—it appears in the capability sections and the source list instead, never with an invented price. The orange line joins every offer that set a new highest list price, in date order; it is a price record only, computed from the price column, with no brand or tier name consulted. The dashed lime path connects the first offer to reach GPT-5’s index score to the cheapest offer that has since matched or beaten it. Nothing here forecasts future models or establishes that spending caused these gains.

Launch prices, and dates are not interchangeable. Launch price. Where a model’s price later changed, the chart keeps the launch figure and the change is recorded in laterPriceChange, not on the chart face. The oldest sub-ceiling value is the May 30, 2024 Gemini paid offer already available at the June 30 chart boundary. Its original tariff is corroborated by Google’s later official before/after price table; a single May 30 tariff receipt was not retrieved. The final point uses GPT-5.6 Luna’s July 30 price cut, not its July 9 release date. GPT-5.3 Codex remains in the source catalog but is not plotted: its only verified price is current, not launch-dated. Historical offers may be retired.

Price range is not constant-capability cost. The most expensive offer on the chart has set 6 successive highs, from $7.50/M to $67.50/M—9.0×—and nothing has priced above $67.50/M since Mar 5, 2026. A rising maximum sampled price does not disprove capability deflation: it is a different measure, not a like-for-like comparison. The retained priced comparators—GPT-6 Astra and Claude Fable 5.1 at $10 input / $50 output, $20/M blended—sit below that high, not above it. Their 90% confidence intervals overlap; a higher point estimate does not establish a statistically decisive lead.

GPT-4o (May 2024)May 13, 2024 · 128.99 ECI · 90% interval 124.62–131.23

$5/M input + $15/M output = $7.5/M blended. Launch price of the pinned gpt-4o-2024-05-13 snapshot, $5 input / $15 output, read from the archived OpenAI pricing page and unchanged on the live page at the September 21, 2026 check.

Alias: The pinned gpt-4o-2024-05-13 id never changed price. The unpinned gpt-4o alias now resolves to $2.50 / $10; we plot the pinned id that matches the dated capability row.

Claude 3.5 SonnetJun 20, 2024 · 130.00 ECI · 90% interval not published by the source

$3/M input + $15/M output = $6/M blended. Launch price $3 input / $15 output from Anthropic’s archived model table, five days after release.

Availability: Historical rate. Claude 3.5 Sonnet has been retired from Anthropic’s first-party API and no longer appears on the live pricing table; it remains reachable through cloud resellers.

o1-miniSep 12, 2024 · 135.70 ECI · 90% interval 132.36–137.17

$3/M input + $12/M output = $5.25/M blended. Launch price $3 input / $12 output from the archived OpenAI pricing page two days after release.

o1Dec 17, 2024 · 141.72 ECI · 90% interval 139.71–142.94

$15/M input + $60/M output = $26.25/M blended. Launch price $15 input / $60 output from the archived OpenAI pricing page. Its footnote states output tokens include billed internal reasoning tokens.

o3Apr 16, 2025 · 146.73 ECI · 90% interval 144.83–148.52

$10/M input + $40/M output = $17.5/M blended. Launch price $10 input / $40 output, read from the archived pricing page the day after release.

Later price change, not plotted: OpenAI cut o3 by 80% on June 10, 2025 to $2 input / $8 output ($3.50 blended), where it remains. We plot the launch price, so o3 sits 5× to the right of where its current price would put it.

Gemini 2.5 Pro (Mar 2025)Apr 30, 2025 · 143.96 ECI · 90% interval 141.74–146.84

$1.25/M input + $10/M output = $3.4375/M blended. First paid tariff, $1.25 input / $10 output for prompts at or under 200K, unchanged since. Output pricing explicitly includes thinking tokens.

Caveat: Gemini 2.5 Pro launched March 31, 2025 as a free experimental model. This is its first PAID price, observed about one month later, so the price date is later than the release date by design.

o3-proJun 10, 2025 · 147.33 ECI · 90% interval 145.56–150.10

$20/M input + $80/M output = $35/M blended. Launch-day OpenAI Developers announcement gives $20 input / $80 output; unchanged on the live model page at the September 21, 2026 check.

GPT-5Aug 7, 2025 · 150.00 ECI · 90% interval not published by the source

$1.25/M input + $10/M output = $3.4375/M blended. Launch announcement prices GPT-5 at $1.25 input / $10 output; unchanged since. Same offer as the flagship reference diamond.

GPT-5 ProOct 7, 2025 · 150.59 ECI · 90% interval 148.94–154.18

$15/M input + $120/M output = $41.25/M blended. Launch price $15 input / $120 output, archived two days after release and unchanged on the live page. Responses-API only, with no cached-input rate.

Gemini 3 ProNov 18, 2025 · 153.07 ECI · 90% interval 151.06–155.42

$2/M input + $12/M output = $4.5/M blended. Launch price $2 input / $12 output for prompts at or under 200K, archived two days after release. Output pricing includes thinking tokens.

GPT-5.2 ProDec 11, 2025 · 156.33 ECI · 90% interval 153.89–159.98

$21/M input + $168/M output = $57.75/M blended. $21 input / $168 output. The GPT-5.2 launch post priced only the non-Pro model; the figure comes from OpenAI’s own later pricing table and the live page, which agree.

Caveat: No provider artifact dated December 11, 2025 was retrieved for the Pro tier. Launch-day third-party coverage reported the same $21 / $168, and the provider figures have not changed, so launch and current price appear identical.

GPT-5.3 CodexFeb 5, 2026 · 157.09 ECI · 90% interval 153.56–161.38

$1.75/M input + $14/M output = $4.8125/M blended. $1.75 input / $14 output from the live OpenAI model page.

Caveat: Current price only: no launch-dated provider artifact for GPT-5.3 Codex could be retrieved, so it is retained in the source catalog but excluded from the launch-price scatter.

GPT-5.4 ProMar 5, 2026 · 159.57 ECI · 90% interval 157.10–162.70

$30/M input + $180/M output = $67.5/M blended. Launch-day pricing table gives $30 input / $180 output; unchanged on the live page. Short-context tier: the long-context rate above 272K is $60 / $270.

GPT-5.5 ProApr 24, 2026 · 162.91 ECI · 90% interval 160.04–166.60

$30/M input + $180/M output = $67.5/M blended. Launch price $30 input / $180 output, identical to GPT-5.4 Pro. Short-context tier; the long-context rate above 272K is $60 / $270.

Caveat: GPT-5.5 Pro reached ChatGPT on April 23, 2026 and the API one day later, on April 24. The capability row is dated to the release; the price date is the API date.

Claude Fable 5Jun 9, 2026 · 163.42 ECI · 90% interval 160.53–168.18

$10/M input + $50/M output = $20/M blended. Fable 5 launched June 9 at $10 input/$50 output. Its historical attained-capability record remains on the chart; access was suspended June 12 and restored July 1, so this does not claim continuous API availability. Use public Epoch ECI for Fable 5, not the separately safeguarded Mythos model or Anthropic internal ECI.

Claude Fable 5.1Sep 1, 2026 · 166.49 ECI · 90% interval 163.47–171.35

$10/M input + $50/M output = $20/M blended. September 1 release keeps the uncached $10 input/$50 output tariff. Its cache-read discount does not enter our uncached 75/25 basket. Use public Epoch ECI for Fable 5.1, not Mythos 5.1 or Anthropic internal ECI.

GPT-6 AstraSep 3, 2026 · 168.81 ECI · 90% interval 165.45–174.51

$10/M input + $50/M output = $20/M blended. September 3 GPT-6 Astra release, standard short-context uncached API tariff $10 input/$50 output. Public Epoch ECI reflects the benchmark collection across reasoning settings, not guaranteed quality for a fixed token count or task budget.

Availability: September 3 is the official announcement date, with phased API rollout; this date does not guarantee universal API access.

Gemini 1.5 Flash (May 2024)May 30, 2024 · 122.68 ECI · $0.525 / M

Google opened paid Gemini API billing May 30. Its August 8 official before/after table explicitly preserves the earlier .35/1.05 tariff; the new .075/.30 price begins August 12. Epoch names the original May checkpoint and dates its release May 23. This joins the original paid launch to the preserved pre-cut tariff, not the August price to the May release.

GPT-4o miniJul 18, 2024 · 126.46 ECI · $0.2625 / M

Dated release identifies the API model and .15 input/.60 output rates. ECI row is GPT-4o mini, not the larger GPT-4o.

Gemini 1.5 Flash (Sep 2024)Sep 24, 2024 · 129.33 ECI · $0.1313 / M

September24 launch explicitly identifies gemini-1.5-flash-002; the already-effective August12 Flash tariff is .075/.30 for prompts below128K. The much less capable May model is not assigned the September ECI score.

DeepSeek-R1Jan 20, 2025 · 138.82 ECI · $0.96 / M

Original January R1 announcement explicitly states cache-miss .55 input and2.19 output. Do not use later R1-0528 score or current deepseek-reasoner alias.

Gemini 2.5 Flash (May 2025)May 20, 2025 · 141.43 ECI · $0.9875 / M

Official changelog identifies the 05-20 preview on May 20. The June 17 announcement preserves its previous .15/3.50 thinking tariff and states GA has no model changes from 05-20. Join the exact May checkpoint in the pinned Epoch vintage, not the June row, and use thinking output pricing rather than the cheaper non-thinking tariff.

GPT-5 miniAug 7, 2025 · 145.34 ECI · $0.6875 / M

Dated announcement lists GPT-5 mini .25 input/2 output. ECI row is the mini model, not GPT-5 or GPT-5 chat.

MiniMax-M2.5Feb 12, 2026 · 146.59 ECI · $0.825 / M

Use explicitly priced Lightning serving at .30/2.40. Provider states M2.5 andM2.5-Lightning are identical in capability and differ in speed; this supports the same M2.5 ECI. Do not infer the ambiguous half-price standard input rate.

GPT-5.6 LunaJul 30, 2026 · 156.50 ECI · $0.45 / M

The July 30 announcement gives effective .20/1.20 rates. The model release is July 9; this milestone is dated July 30 because the later cut cannot be backdated to release. It is a serving-efficiency price cut, not a newly substituted model score.

Capability-index dataset and methodology ↗
BEYOND THE CHECKBOOKS / PUBLIC DATA

More capable. More widely used.

Two independent checks on the AI story: harder tasks and businesses actually paying.

CAPABILITY / METR TASK HORIZONS

From minutes to hours.

3.1 hourshuman-expert task time
at 80% predicted AI success
0h2h4h1.7 min3.1 hoursLAST MODEL: APR ’26NO EXTRAPOLATIONJun 2024Jun 2025Aug 202680% success · release-date record · zero-based task-hours axis
THEBRRR READ

From tiny tasks to multi-hour work.

111×task horizon · Jun ’24–Apr ’26
3.1hhuman-expert task difficulty
80%predicted success threshold

At the same 80% predicted success rate, METR’s task-horizon estimate rose from 1.7 minutes to 3.1 hours. That’s the time a human expert needs for the task—not how long an AI runs unattended.

The opportunity is moving beyond quick answers toward more substantial software work.

Point estimates, not 111× productivity. Latest early-preview result: 95% interval 1.6–6.6 human-expert hours.

Explore task estimates & uncertaintyData + sources ↗
Apr 2026 · Claude Mythos Preview (early)3.10 hr · 95% interval 1.62 hr–6.64 hr
18 models · METR TH1.1

Task difficulty measured by human-expert completion time—not how long an AI runs unattended. Software and reasoning tasks; not a general workplace productivity measure. Latest model is an early preview.

One evaluation vintage. No cost comparison.

All results use METR TH1.1, updated May 8, 2026, applied retrospectively to release dates. Different agent scaffolds can affect results. Dots show tested models; the line connects record-setting point estimates. Small differences may fall within uncertainty. No claim that these models meet the adjacent hero’s $1 token-price ceiling.

METR · snapshot updated May 8, 2026
DEMAND / RAMP PAID AI ADOPTION

Are businesses paying for it?

55.9%of Ramp-observed businesses
+0.6 pp in Sep 2026
0%20%40%60%31.4%55.9%Jun 2024Jun 2025Sep 2026Monthly observations · zero-based percentage axis
THEBRRR READ

From one in three to more than half.

55.9%Ramp-observed paid adoption
+24.5ppsince June 2024
+0.6pplatest month · Sep 2026

Ramp’s paid AI share rose 24.5 percentage points—from 31.4% to 55.9% since June 2024. More than half now show a payment to an identified AI product or service—not just free experimentation.

A broader base of paying businesses strengthens the demand side of the AI buildout.

Ramp-observed businesses, not the whole economy. Revised source history. Top-1% spending was 3.1% below July in September.

Explore adoption & spending intensityData + sources ↗
Sep 202655.86% paid adoption

Payments observed through Ramp—not all U.S. businesses, revenue growth, or proof of returns on the buildout.

Breadth is growing. Spending intensity can still fall.

Ramp’s October report says median AI spend per employee among its top 1% of firms reached $9,568 in September, still 3.1% below July. This small group is volatile. The current source revises history after adding vendors and later transactions; the chart uses one complete revised vintage.

Read the October report
Ramp AI Index · through Sep 2026
What’s underneath: ~3.4× annual growth in AI compute stock. Epoch’s published estimate since 2022; 90% interval 3.2–3.7×. Estimated chip stock—not utilization, a current run rate, or a forecast. Source last updated Feb 5; unchanged on Oct 26 review.Epoch’s data & methods ↗
KEEP EXPLORINGArtificial Analysis · current model economicsOpenRouter · routed usage, not total AI demandEpoch · physical clusters
Earlier chapter: the collapse in token prices2023–2025 · historical offers
02 / THE BUYING-POWER LEAP

The 214× leap in AI buying power.

A fixed benchmark bar. A fixed budget. Far more tokens.

571M tokens / $100versus 2.7M at the first selected offer · MMLU ≥86%
HISTORICAL MILESTONE
FEB 2025214×buying power
0M200M400M600M2.7MMar 20236.7MNov 202313MMay 2024571MFeb 2025Million blended tokens per $100 · 75% input / 25% output · zero-based linear scale
0M300M600M2.7M571MMar 2023Feb 2025Selected offers. Dashed lines are visual guides.
Selected verified price offersTap or focus a point for its model
Gemini 2.0 Flash · Feb 2025571.4M tokens for $100 · $0.175 per 1M tokens
HISTORICAL SERIES · NOT TODAY’S PRICE FRONTIEROpen this price announcement

More tokens at a benchmark threshold—not 214× more intelligence or solved tasks. Selected offers end February 2025; no later trajectory is invented. Benchmark evaluation methods may differ.

CAPABILITY CHECKPOINT

Reasoning scores. Measured task costs.

Latest source snapshot; retained Astra/Fable comparators below, not a claim that these are the newest models.

SOURCE SNAPSHOT Checked
Semi-Private evaluation · benchmarks kept separate
MAX REASONING

GPT-6 Astra

Announced
95%ARC-AGI-2 score
$1.12per tested task
MAX REASONING

Claude Fable 5.1

Announced
90%ARC-AGI-2 score
$4.49per tested task
MAX REASONING

Claude Fable 5

Announced
89.2%ARC-AGI-2 score
$5.45per tested task

Same Semi-Private split. Each model’s Max setting—not equal compute budgets or necessarily its best score. Published retail test costs, including unsuccessful tasks.

How cheaply can you clear the bar?

Choose a score floor. The lowest-cost qualifying model may be a different one.

LOWEST REPORTED COST AT ≥75%7.67¢US cents per tested task
MEASURED ARC-AGI-2 SCORE76.8%
≥75% floor
Dots3-Note Preview (Max)
SET YOUR MINIMUM SCORELowest reported cost / tested task

Published test costs, not a live API quote or availability guarantee. This is a quality–cost snapshot, not a historical trend.

What these numbers mean

ARC tests novel visual puzzle solving. The score is benchmark-specific, not a percentage of general intelligence. Cost is per tested task, including unsuccessful tasks—not the price of a successful real-world outcome.

For each score floor, we select the lowest positive test cost among the source’s displayed base-language-model and chain-of-thought evaluations on the same benchmark and Semi-Private split. Custom systems are excluded. Model checkpoint and reasoning effort stay attached to each result.

The featured comparison fixes the displayed reasoning setting at Max for GPT-6 Astra, Claude Fable 5.1 and Claude Fable 5. Reasoning labels are provider-specific, so this is not a matched-compute experiment. Other settings can be cheaper or score higher; the score-floor selector independently searches all eligible reported settings.

ARC lists September 2 for Astra’s model date; OpenAI announced it September 3 with phased rollout. Cards show the announcement date; the downloadable data preserves ARC’s original date. The reason for the discrepancy is not verified.

The source dataset was generated on Oct 21, 2026 and checked Oct 26, 2026. Its date is not a model’s launch date or a verified historical price date. ARC-AGI-1 and ARC-AGI-2 results are never combined; these figures do not measure electricity use or the return on infrastructure spending.

Selected source value: $0.07665574 USD per tested task; score 76.8%.

Source: ARC PrizeLeaderboardSource dataMethodology
LOOK AHEAD / KNOW WHAT YOU’RE LOOKING AT

Plans where we have them.
Scenarios where we don’t.

The dashed path preserves the June-quarter-vintage model: Alphabet guidance plus June cash-spending pace. Meta has since reported Q3 cash PP&E of $33.942B (model: $30.116B). The other three Q3 reports are not yet available. This is a frozen what-if—not an updated forecast.

SLOWER PACE$644BModeled TTM at Jun 2027
REFERENCE PATH$719BModeled TTM at Jun 2027
FASTER PACE$802BModeled TTM at Jun 2027

The key insight: even a flat modeled quarterly pace lifts the annual total as older, lower-spending quarters roll off. A rising TTM line does not mean quarterly spending keeps accelerating.

Alphabet · 2026-07-22

$195–205B capex in calendar 2026. Significant increase expected in 2027; no dollar amount disclosed.

Comparable cash basis. H2 is guidance minus $80.598B H1 actuals; equal quarterly allocation is our assumption.

Official company source

Amazon · 2026-07-30

Approximately $220B cash capex in 2026, increased from about $200B. Company cash-capex convention is net of sales/incentives.

Not directly spliced into gross cash PP&E. Model holds Q2 reported purchases at $54.208B per quarter before pace sensitivities.

Official earnings audio · 13:10

Microsoft · 2026-07-29

About $175B calendar-2026 capex; fiscal Q1 2027 above $50B. FY2027 capex expected to grow.

Includes finance leases; cash payment timing also differs. The $190B-to-$175B change is classification, not an investment cut. Model holds $35.802B quarterly cash PP&E.

Official company source

Meta · 2026-10-21

$135–145B calendar-2026 capex, including finance-lease principal; narrowed from $130–145B.

Not comparable to cash PP&E alone. Meta reported Q3 cash PP&E $33.942B versus the frozen model $30.116B. No unreported lease bridge is invented.

Official company source
Open every assumption and calculation
  • Alphabet: 2026 cash capex guidance $195–205B minus reported H1 cash PP&E $80.598B gives $114.402–124.402B for H2. Split equally between Q3 and Q4 for this illustration; management did not give that quarterly schedule.
  • Frozen reference path: Alphabet uses $200B annual guidance midpoint and $59.701B per modeled quarter. Amazon, Microsoft and Meta hold their June-quarter cash PP&E pace: $54.208B, $35.802B and $30.116B. Meta subsequently reported Q3 cash PP&E of $33.942B; this historical-vintage sensitivity is not an updated forecast.
  • Only $119.402B (16.6%) of the $719.308B reference next-four-quarter total is directly derived from a comparable management guidance range. The rest is pacing assumptions; no 2027 management dollar target is inferred.
  • Slower/faster sensitivities: Amazon, Microsoft and Meta change by −5%/+5% sequentially each future quarter. Alphabet starts at its 2026 guidance lower/upper endpoint for H2, then changes that H2 quarterly pace by −5%/+5% each quarter in H1 2027.
  • The −5%/+5% rates are editorial stress assumptions, not statistical confidence bounds, worst/best cases, or probabilities. Actual outcomes can fall outside the band.
  • The reference aggregate quarterly spend is flat at $179.827B in each future quarter. Its TTM line rises as older, smaller quarters roll off; it does not imply continued quarterly acceleration.
  • Dashed segments remain the frozen model, including Meta Q3 which has since been reported separately. No four-company Q3 actual aggregate exists yet. Lease-inclusive management guidance is context, not mechanically added to cash-only actuals.
  • The outlook covers Q3 2026–Q2 2027: twelve months after the June 2026 reporting cutoff, not twelve months from the September 2026 research date.

Download the model inputs and all quarterly scenarios →

THE LATEST PRINT / COMPANY BY COMPANY

Who’s stepping on the accelerator?

Quarterly cash purchases · identical chart scales
Latest reported quarter: Q2 2026

Amazon

2026 Q2 · reported cash PP&E

$54.2B
Same scale · $0–60B+68% YoY

Alphabet

2026 Q2 · reported cash PP&E

$44.9B
Same scale · $0–60B+100% YoY

Microsoft

2026 Q2 · reported cash PP&E

$35.8B
Same scale · $0–60B+110% YoY

Meta

2026 Q2 · reported cash PP&E

$30.1B
Same scale · $0–60B+82% YoY
FOLLOW THE CHECKBOOKS

Four giants. $294.8B put to work.

January–June actuals, compared with January–June 2025 ($160.1B). Not an annualized forecast. Company accounting presentation differs; see methodology.

THE INVESTOR LENS / FOLLOW THE CASH

Can the cash engine
fund the next leg?

The buildout has real businesses behind it.
Here’s how hard their cash is working.

12 MONTHS ENDED JUN 2026 · REPORTED
$660.3Bgenerated from operations

For every $100 generated,
$77 went into cash PP&E.

$510.7Bcash PP&E · reported
$149.6Boperating cash minus PP&E

The cash engine grew +33.9% in a year. Spending grew +75.3%. Reinvestment climbed from $59 to $77 per $100 of operating cash.

A formidable cash engine backs the buildout. The investment test: can cash generation catch up with the spending ramp?

Operating cashCash PP&ESame dollar scale · zero baseline
AmazonAMZN107.2% reinvested
$161.4B
$173.0B

−$11.6B operating cash minus PP&E · spend exceeds cash generated

AlphabetGOOGL71.3% reinvested
$185.7B
$132.4B

$53.3B operating cash minus PP&E

MicrosoftMSFT63.4% reinvested
$182.9B
$115.9B

$67.0B operating cash minus PP&E

MetaMETA68.6% reinvested
$130.3B
$89.3B

$41.0B operating cash minus PP&E

A positive aggregate does not transfer cash between companies. This is company-wide funding capacity—not AI revenue, AI returns or a solvency test.

Check the cash math & sourcesReported vs. modeled ↗

All figures in USD billions, using unrounded USD-million inputs. Reported view compares July 2025–June 2026 with July 2024–June 2025. Microsoft’s June-ending fiscal year matches these windows; Meta is annual plus current half-year minus prior half-year.

Operating cash minus cash PP&E is not a harmonized company-reported free-cash-flow measure. It excludes other cash claims including lease principal, acquisitions, debt repayment, dividends and buybacks. Amazon uses gross PP&E purchases, consistent with the spending hero. Operating cash includes stock-compensation addbacks and working-capital effects.

The stress test multiplies each company’s latest annual operating cash by your growth assumption and subtracts the four-company next-four-quarter spending outlook (low, reference and high paths, shown in the outlook section above). Neither the operating-cash assumption nor the spending path is a forecast of investment returns. Balance-sheet cash and financing are not modeled. The aggregate matching-growth figure does not mean every company individually covers spending.

AMZN 2026Q2 MSFT FY26Q4 GOOGL 2026Q2 GOOGL 2025Q2 META 2025FY META 2026H1 META 2025H1
Download the inputs and derivations →
03 / THE FUNDING SHIFT

Operating cash still covers it.
Watch the gap close.

The chart above asks whether the cash engine can fund the next leg.
This one follows what happens as the answer tightens.

REINVESTMENT · TTM JUN 2026 vs TTM JUN 2025
$77of every $100 of operating cash went back into cash PP&E

A year ago it was $59. 1 of 4 is already spending more than it earns from operations.

+18.3points in one year
$149.6Boperating cash left after PP&E

Across $660.3B of operating cash, these four spent $510.7B on property and equipment. That still leaves a surplus. The direction is what matters: two completed twelve-month windows, +18.3 points apart.

This is observed history, not a projection. We do not put a date on the crossover — we show how far it has already moved.

Prior yearTTM Jun 2026Zero baseline · 100% marked
AmazonAMZN107.2% · past 100

88.9% → 107.2% · +18.3 points

AlphabetGOOGL71.3%

50.1% → 71.3% · +21.2 points

MetaMETA68.6%

51.0% → 68.6% · +17.6 points

MicrosoftMSFT63.4%

47.4% → 63.4% · +16.0 points

Reinvestment is cash PP&E divided by operating cash for the same twelve months. Figures are carried from the cash engine above, so the two sections cannot drift apart.

01 / WHO FUNDS THE REST

Debt issuance more than quadrupled in a year.

$51.3B → $226.2B
  • GOOGL$89.4Bfrom $40.1B
  • AMZN$81.9Bfrom $0.7B
  • META$54.8Bfrom $10.4B
  • MSFT$0.0Bfrom $0.0B

Microsoft is the outlier: zero reported debt-issuance proceeds in both windows. The rest raised $226.2B between them in the latest year, against $51.3B the year before.

Gross proceeds as each company reports them, not net new borrowing. The four do not share one accounting line — see the drilldown.

02 / OFF THE CASH-CAPEX LINE

Finance leases move spend where the capex chart cannot see it.

$22.9B → $30.6B
  • MSFT$24.6Bfrom $20.5B
  • AMZN$4.0Bfrom $1.6B
  • GOOGL$1.9Bfrom $0.8B

New finance-lease commitments are noncash. They never touch the cash purchases line the spending hero tracks. Microsoft alone is 81% of these additions — $24.6B against $115.9B of cash PP&E.

That is the same mechanism behind Microsoft’s July 29, 2026 guidance change. The move from roughly $190B to $175B was a classification change, not a cut — the spending did not leave, it changed which line it lands on.

Covers 3 of 4 companies. Meta discloses right-of-use assets obtained for new finance leases only on an annual basis. No interim six-month figure is filed, so a trailing-twelve-month value ended June 30, 2026 cannot be derived from filings. Omitted rather than estimated. Meta's filed annual values are $0.181B for FY2024 and $0.613B for FY2025.

03 / THE CREDIT BACKDROP

Credit spreads have widened. Keep the scale in view.

0.89%Investment gradewas 0.80% a year earlier
3.40%High yieldwas 3.01% a year earlier

Investment-grade spreads sit at 0.89% and high yield at 3.40%. Against the same date a year earlier that is +0.09pp and +0.39pp. The observed ranges since October 2025 are IG 0.73–0.94% and HY 2.60–3.46%. These market-wide spreads and the $226.2B issuance total have different observation windows; neither establishes a causal link.

ICE BofA option-adjusted spreads via FRED, observed 2026-10-22, compared with 2025-10-22. A dated snapshot, not a live quote. Index-level, not any single company’s cost of funds.

THE DISTINCTION THAT MATTERS

The four can fund this. The companies building for them largely cannot.

CoreWeave is the cleanest filed example, because it discloses its own concentration. In the same twelve months, it generated $6.9B of operating cash and spent $20.6B on property and equipment — 298% of operating cash, against 77% for the four above. It raised $24.2B of debt to do it.

Its contracted backlog is $103.7B, up from $30.1B a year earlier. Its largest customer was 36% of revenue in the quarter, and its top three were 72%. A year earlier a single customer was 71%.

The filing reports customer concentration as unnamed Customer A, B and C. It separately names Microsoft, OpenAI and Meta as significant customers but does not map those names to the lettered rows. This page does not map them either. Percentages are share of quarterly revenue, not of backlog; the filing does not split backlog by customer, so we do not publish one.

298%CoreWeave cash PP&E as a share of its operating cash77%The four hyperscalers, same measure, same window
Data + sources for the funding shiftEvery figure, dated ↗

Windows. Everything compares the twelve months ended June 30, 2026 with the twelve months ended June 30, 2025 — the same windows as the cash engine above. Microsoft’s June fiscal year lines up directly. Alphabet, Amazon, Meta and CoreWeave use full year plus current half minus prior half, except where a filer reports twelve-month June columns outright, which Amazon does.

Retrieval. SEC EDGAR XBRL companyconcept/companyfacts API at data.sec.gov, original retrieval 2026-09-21; FRED fredgraph.csv re-retrieved 2026-10-16, latest observation 2026-10-15. Values are committed to this file and read at build time. Nothing on this page fetches at request time.

Debt issuance. There is no single shared XBRL tag across these four filers. Each company's own reported line is used and named below. Alphabet's line is net of issuance costs and is not restricted to long-term maturities; Microsoft's line covers debt maturing in more than three months. These are gross issuance proceeds, not net new borrowing — refinancing and repayments are not deducted. Per-company derivations: MSFT — ProceedsFromDebtMaturingInMoreThanThreeMonths; Reported directly as fiscal-year columns; both are an affirmatively reported zero, not a missing value. GOOGL — ProceedsFromDebtNetOfIssuanceCosts; Current: FY2025 64564 + H1 2026 56226 − H1 2025 31378 = 89412. Prior: FY2024 13589 + H1 2025 31378 − H1 2024 4875 = 40092. AMZN — ProceedsFromIssuanceOfLongTermDebt; Both reported directly as twelve-month June columns in the Q2 2026 10-Q. No derivation applied. META — ProceedsFromIssuanceOfLongTermDebt; Current: FY2025 29906 + H1 2026 24910 − H1 2025 0 = 54816. Prior: FY2024 10432 + H1 2025 0 − H1 2024 0 = 10432.

Finance leases. Right-of-use assets obtained in exchange for new finance lease liabilities, trailing twelve months. This is a noncash addition: it never appears on the cash purchases-of-property-and-equipment line the spending hero tracks. Tag RightOfUseAssetObtainedInExchangeForFinanceLeaseLiability. MSFT — Reported directly as fiscal-year columns ending June 30. GOOGL — Current: FY2025 1606 + H1 2026 902 − H1 2025 606 = 1902. Prior: FY2024 313 + H1 2025 606 − H1 2024 165 = 754. AMZN — Both reported directly as twelve-month June columns in the Q2 2026 10-Q. Meta discloses right-of-use assets obtained for new finance leases only on an annual basis. No interim six-month figure is filed, so a trailing-twelve-month value ended June 30, 2026 cannot be derived from filings. Omitted rather than estimated. Meta's filed annual values are $0.181B for FY2024 and $0.613B for FY2025.

Credit spreads. Committed dated snapshot. Refreshed by hand when the chapter is revisited; it is not a live quote and does not update per visitor. Observation 2026-10-22; prior-year comparison 2025-10-22; range measured from 2025-10-01.

Concentration. TTM ended June 30, 2026 = FY2025 + H1 2026 − H1 2025. Operating cash: 3058 + 3663 − (−190) = 6911. Cash PP&E: 10309 + 14117 − 3860 = 20566. Debt issuance: 11829 + 16747 − 4396 = 24180. Revenue percentages and RPO are as reported for the stated periods; percentages are three-months-ended figures and are not annualized.

  • Debt issuance is gross proceeds as each company reports it. It is not net new borrowing: repayments, maturities and refinancing are not deducted, and a large refinancing year can raise the figure without raising outstanding debt.
  • The four companies do not share one debt-issuance tag. Alphabet's line is net of issuance costs and not limited to long-term maturities; Microsoft's is debt maturing in more than three months; Amazon's and Meta's are long-term debt issuance. Do not read the four as a harmonized series.
  • Finance-lease additions are noncash and cover three of four companies. Meta files this disclosure annually only, so it is excluded from the trailing-twelve-month aggregate rather than estimated.
  • Finance-lease additions are commitments recognized at inception, not cash spent in the window. They are not additive to cash PP&E as a single spending figure.
  • Credit spreads are index-level option-adjusted spreads for the whole US corporate and high-yield markets. They are not the borrowing cost of any company on this page and not an AI-specific credit measure.
  • CoreWeave is one levered AI-infrastructure operator, chosen because its concentration is disclosed in its filings. It is not a representative sample of the sector and nothing here is a claim about other operators.
  • Customer percentages are share of revenue for a three-month period, not share of remaining performance obligations. The filing does not disclose an RPO split by customer, so no backlog-by-counterparty figure is presented.
  • Reinvestment above 100% is an accounting comparison against one window's operating cash, not a solvency test or a funding gap. Existing balance-sheet cash, debt and equity access can finance a difference.
  • Every figure on this page is a committed dated observation read at build time. None of it updates on its own.
MSFT FY2026 10K MSFT GUIDANCE FY26Q4 GOOGL 2026Q2 10Q GOOGL 2025FY 10K AMZN 2026Q2 10Q AMZN 2025FY 10K META 2026Q2 10Q META 2025FY 10K CRWV 2026Q2 10Q CRWV 2025FY 10K FRED BAMLC0A0CM FRED BAMLH0A0HYM2
THE NEXT INSTRUMENT / THE AI TRILLION-DOLLAR DERBY

Same giants. Different race.

Switch from market value to quarterly sales. Watch the field reorder—and see the distance between the leaders.

Enter the derby
WHAT THIS DOES—AND DOESN’T—TELL US

Efficiency is not the opposite of a buildout.

THE QUESTION FOR INVESTORS

Cheaper can mean more demand.

Lower prices can make new uses economical. The open question is whether usage grows fast enough to support the spending—and which companies capture the value. These charts frame that question; they don’t prove the answer.

WHY DOLLARS, NOT ELECTRICITY?

Measure what we can defend.

For our historical buying-power chart, API prices have dated, public records. A comparable “intelligence per joule” history needs consistent tasks, model quality, hardware, and utilization. We haven’t established that series, so we don’t convert token prices into power efficiency.

CHECK OUR WORK

Big numbers. Open books.

Sources checked October 26, 2026 (common June financial window; Meta Q3 separately reported; frozen scenario retained). A sourced snapshot, not a live feed.

Download data & sources
How do we measure the buildout?

We add reported cash purchases/additions of property and equipment, aligned to calendar periods. The spending ramp covers five companies — Alphabet, Amazon, Meta, Microsoft and Oracle. The critical signals, company breakdown, outlook and funding sections cover the original four, without Oracle. Values are USD billions. This financial proxy includes non-AI assets; it is not a direct count of chips, energized data centers, AI-only spending, or training costs.

Lease principal and noncash lease additions are excluded. Amazon’s figure precedes sales/incentive offsets; some historical Meta statements label purchases as net. The figures retain these disclosure differences rather than claiming perfect accounting harmonization. Microsoft’s fiscal statements are reconstructed into calendar quarters. Oracle’s fiscal quarters end one month before calendar quarters, so its “calendar year” runs December through November; this is the closest alignment its filings allow.

The spending ramp shows one bar per calendar year: 2019–2025 as reported, and 2026 as an estimate (the four’s reported first half plus the outlook reference path; Oracle’s three reported quarters plus its fiscal-2027 capex guidance for the last one). The 2027E bar adds TheBRRR house estimates: frozen 2026 cash baselines × growth (Alphabet 30%, Amazon 25%, Meta 35%, Microsoft 25%, Oracle 20%), rounded to $1B each; not guidance or consensus. See the visible Chart01 caveat and model ledger. Its 5.8× headline compares 2025 with 2019 and never uses the estimate. Elsewhere, the latest four-company 12-month value adds Q3 2025 through Q2 2026 actuals. The outlook section’s reference path is flat at $179.827B per future quarter; the trailing annual total rises as smaller historical quarters roll off. YoY compares equal-length periods. Four-company annual totals: 2023 $147.2B → 2024 $228.3B → 2025 $376.1B. Derived values use unrounded figures.

Download Chart01 annual CSV Download quarterly CSV
What does “214× buying power” actually mean?

We compare four selected first-party API offers whose source-reported MMLU scores meet or exceed 86%. MMLU measures performance across academic knowledge tasks; it is not a universal unit of intelligence. Scores come from Epoch AI’s compiled dataset, not our own evaluation. Methods and model snapshots may differ.

Each blended price is 0.75 × input price + 0.25 × output price per million tokens. Thus $37.50 ÷ $0.175 ≈214.3. For a fixed $100 budget, those historical prices buy about 2.7M versus 571.4M blended tokens. This is buying power at a benchmark threshold, not cost per solved task, business ROI, or 214× more intelligence. It does not establish current availability of these historical offers. Standard text rates exclude caching, batch discounts, and free tiers. Launch access was sometimes preview or waitlisted.

Dates are backed by price announcements. We do not attach later discounts to earlier launch dates. Selected observations end February 5, 2025; no continuation into 2026 is implied. Connecting lines are visual guides, not measured intermediate prices or a claim to cover every cheaper model.

Open the primary spending sources
Full quarter-by-quarter derivations and source map →
Can I reuse these charts?

Yes. Download a portrait or landscape graphic from “Share the big picture.” Keep the dates, definitions, sources, and TheBRRR credit visible. Link back here so your readers can check the numbers.

The numbers are free. Get the perspective.Get the research