Company-wide cash PP&E, not AI-only. Outlook is TheBRRR’s model—not company guidance or a confidence interval. Horizon is 12 months after the last reported quarter.
02 / THE INTELLIGENCE FRONTIER
The frontier leaps. Capability gets cheaper.
+33.7
CAPABILITY INDEX POINTS AT ≤$1 / MILLION TOKENS
FRONTIER · SEP 3, 2026GPT-6 Astra166.3 ECI
Record under $1/MAll-price capability record
THEBRRR READ
New ceilings. Lower entry costs.
166.31GPT-6 Astra · ECI
164.47Claude Fable 5.1 · ECI
−87%Luna token price vs. GPT-5 launch
GPT-6 Astra and Fable 5.1 lift the frontier to 166.31 and 164.47 ECI, at $20/M blended tokens each. Meanwhile, GPT-5.6 Luna scores 156.31—above GPT-5’s 150—for just $0.45/M, or 7.6× the tokens per dollar versus GPT-5’s launch price.
The ceiling is rising while yesterday’s frontier becomes affordable to deploy.
Epoch AI · Sep 14, 2026. One index vintage; overlapping 90% intervals. Points, not “% smarter.” Token prices are not task costs.
SHOW CONTEXT
Explore the model milestones & pricesData + sources ↗
A historical record from eight selected 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.
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 same next-four-quarter spending model used 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.
Enter the derby What “smarter under $1” measuresMethod + benchmark and price sources
A fixed price ceiling, not a fixed task budget. Each point is a dated API offer costing no more than $1 per million tokens at a 75% uncached text input / 25% billed output mix, including reasoning tokens. Different models can use different numbers of tokens to solve a task. These are short-context, standard on-demand rates; no free tiers, caching or batch discounts.
One capability ruler.Epoch Capabilities Index combines benchmark results. All scores use the same Sep 14, 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.
What the steps mean. The bright line records the best capability attained in our eight-offer catalog as of each offer date; it does not assert every model remained available between observations. The gray line is the highest score in Epoch’s broader 266-model dataset, regardless of price or access restrictions. Neither line forecasts future models or establishes that spending caused these gains.
Dates are not interchangeable. The first 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. Historical offers may be retired.
The new premium frontier. GPT-6 Astra and Claude Fable 5.1 use standard uncached short-context rates of $10 input / $50 output per million tokens: $20 at the same 75/25 mix. They are not members of the under-$1 series. Their 90% confidence intervals overlap; a higher point estimate does not establish a statistically decisive lead.
$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.
$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.
$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.
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.
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.
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.
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.
MiniMax-M2.5Feb 12, 2026 · 146.51 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.
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.
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
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.
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.
56.1%of Ramp-observed businesses +0.4 pp in August 2026
THEBRRR READ
From one in three to more than half.
56.1%Ramp-observed paid adoption
+22.7ppsince June 2024
+0.4pplatest month · August 2026
Ramp’s paid AI share rose 22.7 percentage points—from 33.5% to 56.1% 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. Adoption is not revenue; top-1% spend per employee fell 9.7% in August.
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 September report says median AI spend per employee among its top 1% of firms fell 9.7% in August. This small group is volatile; July was revised as more transactions arrived.
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 September run rate, or a forecast.Epoch’s data & methods ↗
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
GPT-6 and Fable, side by side.
Frontier reasoning. The scores, the settings, and what the test actually cost.
SOURCE SNAPSHOT Checked
Semi-Private evaluation · benchmarks kept separate
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%24.94¢US cents per tested task
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 Sep 4, 2026 and checked Sep 14, 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: $0.24944952 USD per tested task; score 84.6%.
Plans where we have them. Scenarios where we don’t.
The dashed path combines Alphabet’s comparable 2026 guidance with the other companies’ latest cash-spending pace. It is a what-if—not a four-company 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.
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.
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.
Reference path: Alphabet uses the $200B annual guidance midpoint and $59.701B per future quarter. Amazon, Microsoft and Meta each hold their Q2 2026 reported cash PP&E pace: $54.208B, $35.802B and $30.116B per quarter.
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.
All future chart segments are modeled. Amazon net-cash capex, Microsoft lease-inclusive capex and Meta capex including lease principal are displayed as 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.
January–June actuals, compared with January–June 2025 ($160.1B). Not an annualized forecast. Company accounting presentation differs; see methodology.
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.
THEBRRR / CONNECT THE DOTS
The curves are moving. Know what they mean.
Get the AI and macro research behind the numbers. Who benefits. What could break. What to watch next.
Free to read Source-led research Unsubscribe anytime
CHECK OUR WORK
Big numbers. Open books.
Sources checked September 13, 2026. A sourced snapshot, not a live feed.
We add reported cash purchases/additions of property and equipment for Amazon, Alphabet, Microsoft, and Meta, aligned to calendar periods. 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.
The latest 12-month value adds Q3 2025 through Q2 2026 actuals. The solid hero line covers Q2 2024–Q2 2026 TTM observations, and the dashed model continues to Q2 2027. The 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. Headline annual totals: 2023 $147.2B → 2024 $228.3B → 2025 $376.1B. Derived values use unrounded figures.
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
Amazon
Gross cash purchases of property and equipment; not net of sales/incentives. Entire Amazon, not AWS-only or AI-only.
CALENDARIZED, not fiscal year headline capex. CY2024 = FY2024 44477 minus Jul-Dec2023 19652 plus Jul-Dec2024 30727. CY2025 = FY2025 64551 minus Jul-Dec2024 30727 plus Jul-Dec2025 49270. Excludes noncash finance-lease additions and lease principal. CY2023 = FY2023 28107 minus Jul-Dec2022 12557 plus Jul-Dec2023 19652 =35202.
Cash purchases of property and equipment only, excludes principal payments on finance leases. Therefore not management headline $72.215B 2025/$39.225B 2024 capex.
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.