Anthropic spent $9.74 billion on research and development in the first half of 2026, 5.3 times the $1.85 billion it spent a year earlier, according to its Form S-1. Over a similar span, METR says the rate at which Anthropic’s models improve on its benchmark index rose about 1.9 times.
By our arithmetic, that means progress per R&D dollar fell by about 63% between the two halves, even as progress per year nearly doubled. Michael Wiebe made the same calculation in an X post: “So progress per R&D dollar fell by 63%, even while progress per year doubled”.
Business Pill · Capex and Depreciation Explained
An evergreen Business Engineer explainer on AI capex and why chip lifetime drives AI profits. It teaches the general idea only and does not discuss Anthropic’s S-1 or METR.
The key insight: Anthropic’s S-1 shows R&D of $9.74 billion in the first half of 2026 against $1.85 billion a year earlier, while METR says the rate of improvement on Anthropic’s benchmark index rose about 1.9 times; by our arithmetic, points per R&D dollar fell about 63%.
Anthropic’s S-1, filed on October 26, has a table of operating expenses by function. It shows research and development of $2,275 million in 2024, $6,338 million in 2025, $1,850 million in the six months to June 30, 2025 and $9,741 million in the six months to June 30, 2026. The filing says R&D mainly consists of compute and infrastructure costs for model development, research and internal usage, plus personnel costs.
METR’s assessment, published on November 5 after Anthropic gave it access to internal data, measures progress with AECI, “Anthropic’s version of the Epoch Capabilities Index”. METR says AECI scores “have accelerated 1.9x [95% confidence interval: 1.5–2.8x] since October 2025”, and its chart shows a trend of “+15 points per year before October 2025” that “steepens to +29 points per year after it”.
Anthropic’s R&D expense rose 5.3x between the first halves of 2025 and 2026 while METR’s fitted AECI trend rose 1.9x (+15 to +29 points a year). Data: Anthropic Form S-1 (26 Oct 2026); METR (5 Nov 2026). Indexing is our arithmetic.
What Was Posted on X
good thing none of the models are benchmaxxed, right? 😬
A post by @CompleteSkeptic at 01:04 UTC on 8 November that quotes @sebkrier’s chart of METR’s fitted AECI trend against Anthropic’s R&D spending, with the line “good thing none of the models are benchmaxxed, right?”. That is the poster’s view; neither METR nor Anthropic says the models were tuned to the benchmarks, and METR lists possible overfitting to the types of tasks in AECI as one of its own caveats.
Anthropic’s measured AI capability gains accelerated, but R&D spending grew faster, according to an analysis of METR’s data.
Progress per dollar fell 63% between the first halves of 2025 and 2026, even as annual progress nearly doubled.
A post by @kimmonismus at 10:08 UTC on 8 November summarising the same calculation: citing an analysis of METR’s data, it says progress per dollar fell 63% between the first halves of 2025 and 2026, even as annual progress nearly doubled.
Progress per Dollar
Applying METR’s fitted rates to each half-year gives 7.5 AECI points in the first half of 2025 and 14.5 in the first half of 2026. Divided by R&D spend, that is about 4.05 points per billion dollars in the first half of 2025 and about 1.49 in the first half of 2026, by our arithmetic. Those match the figures in Wiebe’s post.
A chart posted on X by @sebkrier draws the same trend against cumulative R&D spending, with the calculation credited to Claude. Its note reads: “Descriptive only: falling points per dollar does not show that AI failed to raise research productivity”. Another X user, quoting that chart, wrote “good thing none of the models are benchmaxxed, right?”. That is the poster’s view. METR lists a related caveat itself: “recent progress may be overfitting to the types of tasks which are included in AECI”.
What METR Says About Its Own Number
METR says its robustness checks covered “benchmark inclusion, model dating, model inclusion, and benchmark saturation”, and that they “find acceleration rates between 1.1x and 2.9x”. It lists reasons it could be overestimating the acceleration, including “AECI’s focus on verifiable tasks”, and reasons it could be underestimating it.
R&D rose from $1.85 billion to $9.74 billion between the two first halves, by the S-1’s own figures, while the benchmark trend METR measured roughly doubled its speed. By our arithmetic, spending grew about 2.7 times faster than the measured rate of improvement over that period.
The Structural Read
As we read the two documents together, faster progress and costlier progress are both true at once: the measured rate of improvement nearly doubled while the spending behind it grew more than five times.
As we read METR’s assessment, its own caveats matter for any per-dollar reading: it says recent progress may be overfitting to the kinds of tasks in AECI, and its robustness checks give acceleration rates between 1.1x and 2.9x.
As we read the S-1, R&D covers compute, infrastructure and personnel for model development, research and internal usage, so the per-dollar ratio divides a benchmark trend by spending that also pays for other work.
Michael Wiebe, X post, 7 Nov 2026
“So progress per R&D dollar fell by 63%, even while progress per year doubled”
Three Implications
FOR INVESTORS READING THE S-1 The filing’s R&D line rose from $1.85 billion to $9.74 billion between the first halves of 2025 and 2026.
FOR READERS OF METR’S FINDING METR says AECI scores have accelerated 1.9x since October 2025, with a 95% confidence interval of 1.5x to 2.8x.
FOR ANYONE CITING THE 63% The figure is our arithmetic and Wiebe’s; neither METR nor Anthropic reports progress per dollar.
The essay says AI shows “capex commitments that exceed the unit economics of any single application” alongside “real productivity gains”, and adds: “Both are true. The mistake is treating them as competing claims.”
As we read it, the S-1 and METR numbers sit on both sides of that line: spending rising faster than the measured rate of progress, and that rate still rising. The essay does not discuss METR’s assessment or the S-1.
What Is Not Established
METR’s trend is a fitted line on benchmark scores, not a cost measure, and R&D spending also pays for work that does not show up in AECI. The two are not causally linked in either document, and the ratio here is descriptive. Neither METR nor Anthropic reports progress per dollar.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Anthropic’s R&D spend rose 5.3 times between the first halves of 2025 and 2026, according to its S-1, while METR’s fitted benchmark trend rose about 1.9 times. By our arithmetic, AECI points per billion dollars of R&D fell from about 4.05 to about 1.49, roughly 63% lower.
A note on sourcing. Spending figures are from Anthropic’s Form S-1 filed on October 26, 2026. The progress trend is from METR’s assessment of November 5, 2026. Ratios and per-dollar figures are our arithmetic and are descriptive only.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
