# Is AI Speeding Up AI Research at Anthropic?
> Yes. METR's November 2026 assessment finds Anthropic improving its models about 2x faster than before, roughly 80% due to AI and not to more training compute.
Source: https://capitalandcompute.net/blog/is-ai-speeding-up-ai-research-at-anthropic/
Published: 2026-11-07
Updated: 2026-11-07
Yes. METR, the independent AI evaluator, published an assessment on November 5, 2026 that finds Anthropic is likely improving its models on well-specified tasks at about twice the rate it did from March 2024 to October 2025\. METR puts roughly 80% of that speed-up down to AI itself, and it is confident the cause is not a faster scale-up of training compute.
That last point is the one with money attached. Anthropic’s IPO filing says its staff now consume more Claude, priced at list rates, than the company pays them in cash wages. If bigger training runs did not cause the acceleration, the obvious candidate is that internal Claude use: agents writing code and running experiments for Anthropic’s researchers. METR’s report is the first outside estimate of what that spend buys.
Key takeaways
* METR estimates Anthropic is improving model capabilities about 2x faster than in March 2024 to October 2025, and would not be surprised by 1.3x or 3x.
* On its main benchmark index the estimate is 1.9x (95% interval 1.5x to 2.8x): about 15 index points a year before a break in October 2025, about 29 after.
* METR attributes roughly 80% of the current acceleration to AI and rules out an acceleration in training compute as the cause.
* Anthropic redacted most of the evidence METR found most informative and cancelled the compute-verification part of the engagement midway.
* Set against the S-1, more than $186 million a month of list-price internal Claude use is buying about a year of extra progress per year. That is the ROI case, and it cannot be audited from outside.
## What did METR find about Anthropic?
METR (Model Evaluation and Threat Research) ran the engagement from August 12 to October 29, 2026, with 2.5 full-time staff and access to internal data: benchmark scores, algorithmic-efficiency metrics, training compute, lines of code, “AI agent inference spend over time,” self-reported uplift surveys and about 15 staff interviews. It did not get query access to unreleased models. The findings are in [_Independent assessment of AI R&D acceleration at Anthropic_](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/) (METR, 2026).
The core numbers:
* **About 2x.** “Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate” of March 2024 to October 2025\. METR says it would not be surprised if the true figure were 1.3x or 3x.
* **1.9x on the index.** Anthropic’s version of the [Epoch Capabilities Index](https://epoch.ai/eci) (AECI, a composite of many benchmarks) grew about 1.9x faster after an estimated break on October 7, 2025, with a 95% interval of 1.5x to 2.8x. Anthropic’s own Haiku 5.5 system card had put it at 1.7x; METR says the gap comes mainly from adding newer internal-only models.
* **About 80% due to AI.** “We guess that a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI.” METR thinks non-AI factors likely played a bigger part in late 2025.
* **Not compute.** “We are confident that the acceleration was not due to an acceleration in the scaling of training compute.”
* **No near-term wall.** METR found no convincing bottleneck in algorithms, training compute, experiment compute or data likely to bite within 12 months, though it rates its confidence in that view as low.
* **Not autonomous yet.** AI is not making highly significant, fully autonomous contributions to the most important parts of training, but researchers described “signs of life.”
__Anthropic capability index (AECI) trend, March 2024 to October 2026__
| Point | Date | AECI points gained since Mar 2024 |
| ------------------------ | ---------- | --------------------------------- |
| Start of window | Mar 2024 | 0 |
| Estimated breakpoint | 7 Oct 2025 | +24 |
| Fitted trend after break | Oct 2026 | +55 |
| Old pace projected | Oct 2026 | +40 |
Anthropic's capability index rose about 15 points a year until a break on October 7, 2025, then about 29 a year. Projected at the old pace, it would be about 15 points lower by October 2026\. Fitted slopes only; METR's per-model scores are redacted, so the lines are illustrative.Source: [METR, Independent assessment of AI R&D acceleration at Anthropic, November 5, 2026\. Points gained derived by Capital & Compute from METR's published slopes and breakpoint.](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/)
METR describes the jump in plain terms: roughly two years of historical progress in one year. It also tested whether the data fit a one-time jump at Claude Mythos Preview better than a lasting change in slope. The slope change predicted held-out models with an error of 1.9 index points against 3.9 for the jump, so METR treats the faster rate as the new trend rather than a single step.
## How does AI speed up AI research?
The mechanism is ordinary engineering leverage, applied to the work of building models. A frontier lab’s progress comes from three inputs: more training compute, better algorithms (getting more capability per unit of compute), and better data. Training compute grows on a schedule set by data-center deals, which is why METR could check it and rule it out.
Algorithms and data are where AI agents help. Each improvement starts as an experiment: write the code, run a small training job, read the results, change one thing, repeat. If agents write most of that code and run most of those loops, researchers test more ideas per week, and the rate of improvement rises without any extra training compute.
Anthropic’s own measurements describe that shift. Its research arm, the Anthropic Institute, reported in its [measurements for understanding the pace of AI development](https://www.anthropic.com/institute/measuring-pace-of-ai-development) (2026) that Claude “leads” 26% of Anthropic’s AI R&D work as of August 2026, up from under 1% in February, and that more than 90% of the work involves Claude at the level of collaborator or above. That is a prototype index Anthropic built and graded itself, not an independent figure.
## If not compute, what is Anthropic paying for?
Internal inference, and a lot of it. Anthropic’s [Form S-1/A](https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm), filed November 2, 2026, says the list-price value of its own Claude use grew more than 250 times from January to September 2026 and passed monthly cash payroll in April. As worked out in [whether Anthropic’s internal Claude use is worth its payroll](/blog/anthropic-internal-claude-use-vs-payroll/), that bar was about $186 million a month in Q2 2026 and about $204 million in Q3\. Annualized, internal use runs above $2.2 billion at the prices customers pay.
Put the two reports side by side and the ROI case looks like this:
| What | Figure | Source |
| -------------------------------------- | ------------------------------------ | ------------------------------ |
| Internal Claude use, list price | over \~$186M a month from April 2026 | Form S-1/A (derived) |
| Rough compute cost of that use | under \~$90M a month (illustration) | 48 cents compute per revenue $ |
| Capability growth rate vs 2024–25 | about 2x (1.3x to 3x plausible) | METR, November 2026 |
| Share of the speed-up attributed to AI | about 80% | METR, November 2026 |
| Extra progress, October 2025 to 2026 | about one year of old-pace gains | METR slopes, derived |
The cost row is an estimate, not a disclosure. Anthropic spent 48 cents on compute per dollar of revenue in Q2 2026, and that line also pays for training. If internal tokens cost what the average revenue dollar cost to make, $186 million of list-price use was under $90 million of real compute.
On those terms, a lab spending something like $1 billion a year of real compute on internal agents is getting roughly a year of extra capability progress per calendar year. In a race where model releases set pricing power, that is a cheap trade. It also explains why the S-1 frames internal use as something that “compounds these gains.”
What cannot be computed is a precise dollars-per-capability figure. METR saw agent inference spend over time, but those numbers are among the redacted evidence, and Anthropic runs one shared compute pool, so internal use has no separate cost line. The direction is clear; the exchange rate is not public.
## How sure is the 2x estimate?
Less sure than the headline suggests. METR’s own robustness checks spread from 1.1x to 2.9x depending on modelling choices, and most of that spread comes from one technical choice: the function used to turn benchmark scores into a single index.
__METR acceleration estimates for Anthropic by analysis choice__
| Forecaster | Measures | Forecast |
| ------------------------ | ----------------------------------------- | ------------ |
| Headline judgement | METR would not be surprised by 1.3x or 3x | 1.3x to 3x |
| Link function | six ways to combine benchmark scores | 1.1x to 2.9x |
| Snapshot choice | which model checkpoints count | 1.9x to 2.2x |
| Benchmark choice | which benchmarks enter the index | 1.9x to 2.2x |
| Regularization | how hard the fit is smoothed | 1.9x to 2.1x |
| Other labs included | non-Anthropic frontier models added | 1.4x |
| Scaling-policy threshold | Headline figure being checked | 2x |
METR's acceleration estimates ranged from 1.1x to 2.9x across modelling choices. Only the link-function choice moves it far from 1.9x; adding other labs' frontier models lowers it to 1.4x. The dashed line is the 2x level in Anthropic's own scaling policy.Source: [METR, Independent assessment of AI R&D acceleration at Anthropic, robustness checks, November 5, 2026.](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/)
Two other signals point the same way but are weaker:
* **Staff surveys.** In internal Slack polls, average self-reported speed-up from the most capable internal model reached about 12x in October 2026\. METR calls this “only very loose evidence,” and notes that individual uplift is not the same thing as a faster rate of progress.
* **Interviews.** Staff views centred on about 2x, with a wide spread. Of 11 classified answers, 5 were yes or leaning yes, 4 were unclear and 2 were no.
The 1.4x figure deserves weight. When other labs’ frontier models are added, the before-and-after trend is about 17 then 24 points a year. Part of what looks like Anthropic accelerating may be Anthropic catching up from a slow 2023 to 2025 stretch, which METR lists among the reasons it could be overestimating.
Why 2x is the number that matters
Anthropic’s Responsible Scaling Policy sets a 2x acceleration threshold: “dramatic acceleration” of AI progress likely linked to automating AI research. Crossing it triggers stronger safety obligations. A separate METR review says Anthropic’s August 2026 risk report was “overconfident” in saying the threshold had not been crossed.
## What could METR not check?
A lot, and METR says so. Its post states that “Anthropic redacted a majority of the evidence we found most informative for these conclusions,” and that its findings depended significantly on redacted information. Readers are being asked to trust METR’s judgement rather than see the evidence.
Four other limits:
1. **Compute verification was cancelled.** “Anthropic changed the scope of this project midway to exclude compute verification,” including how compute splits between capabilities and safety work. It was dropped before METR reached any conclusion.
2. **No access to unreleased models.** METR could not query them, which is item 1.1 of five items on the AI Evaluator Forum’s minimum-conditions checklist that the engagement did not meet.
3. **Redaction rules tightened late.** Guidance changed on October 27 to 29, adding about five days.
4. **Heavy reliance on Claude.** METR used Claude for data collection, analysis and drafting, and found no significant discrepancies when checking against interviews.
The setup was agreed in advance. Anthropic’s October 15 post, [_An update on independent evaluators_](https://www.anthropic.com/news/an-update-on-independent-evaluators), says it gave six METR staff “employee-like access,” that it “can’t veto publication,” but that it can require redactions within set categories, and that Anthropic decides what falls in them. The redactions were described as removing information “that could accelerate other AI developers.”
The timing is worth stating as fact, without reading intent into it: the redaction guidance tightened in the week before Anthropic filed its amended S-1 on November 2, and the report was submitted the day after.
## Did Anthropic’s own risk report understate the speed-up?
METR thinks so. In a companion piece, [_Review of the AI R&D section of Anthropic’s August 2026 Risk Report_](https://metr.org/blog/2026-11-05-rd-section-anthropic-risk-report-august-2026-review/) (METR, 2026), it agrees catastrophic risk was low as of July 15, but says the public report’s conclusion that the 2x threshold was not crossed was “overstated.”
The private version told a different story. It said internal rates “have less than doubled even if we attributed all of the observed speedup to AI assistance,” and METR worked out the implied figure as 1.97x, a number that did not appear in either version. A reader sees “less than a factor of 2” and assumes well below 2\. METR’s own range was 1.3x to 2.8x, and it “cannot rule out a substantial further acceleration” over the following six months.
For investors reading the S-1, that gap matters in both directions. Faster progress supports the revenue story. It also brings the company closer to scaling-policy obligations that could slow releases.
## What does this mean for teams budgeting AI against headcount?
The lesson is not “spend like Anthropic.” Anthropic pays cost for its own tokens; everyone else pays the rate card, or a negotiated discount from it. The lesson is in how METR measured it.
1. **Measure output growth, not spend.** METR did not ask whether staff felt faster (they reported 12x). It asked whether the rate of improvement changed. Track a delivery metric, such as features shipped or experiments run, before and after the AI budget rises.
2. **Price agent work per task.** Agent loops consume far more tokens than chat. The [Claude Code cost-per-task breakdown](/blog/claude-code-pricing-cost-per-task/) shows how wide that range runs.
3. **Separate the AI share from everything else.** METR’s 80% attribution came from comparing against non-AI explanations, such as new hires, compute and better tooling. Do the same before crediting the budget.
4. **Watch for saturation.** METR notes that Anthropic’s newest model scored above 50% on all 15 benchmarks it was tested on, which makes the next acceleration harder to measure. Internal benchmarks go stale too; refresh them as models improve.
For the lab-level money side, see [how much Anthropic spends on compute](/blog/anthropic-ipo-compute-spending/) and [how Anthropic makes money](/blog/how-does-anthropic-make-money/).
## Frequently asked questions
## Frequently asked questions
Is Anthropic using AI to build better AI?
Yes. METR found in November 2026 that Anthropic is likely improving its models about twice as fast as it did from March 2024 to October 2025, and attributes roughly 80% of that acceleration to AI. Anthropic separately reports Claude leads 26% of its AI R&D work.
What is METR?
METR (Model Evaluation and Threat Research) is a nonprofit that evaluates frontier AI systems, best known for measuring how long a task an AI agent can complete. It received employee-like access at Anthropic under an agreement announced on October 15, 2026.
Did training compute cause the Anthropic acceleration?
No, according to METR, which says it is confident the acceleration was not due to faster scaling of training compute. That points to algorithmic and research-process gains, where internal AI agents do much of the work.
Has Anthropic crossed its 2x AI R&D threshold?
Not officially. Anthropic's August 2026 risk report said no. METR's review calls that conclusion overconfident: its own estimate is about 2x, with a plausible range of 1.3x to 3x, so crossing is plausible.
How much does Anthropic spend on internal AI use?
Its Form S-1/A says internal Claude use at list prices passed monthly cash payroll in April 2026, which was about $186 million a month in Q2 2026\. The real compute cost is lower and is not disclosed.
## Sources
* METR (2026). _Independent assessment of AI R&D acceleration at Anthropic (October 2026)_. METR blog, November 5, 2026\.
* METR (2026). _Review of the AI R&D section of Anthropic’s August 2026 Risk Report_. METR blog, November 5, 2026\.
* Anthropic (2026). _An update on independent evaluators_. Anthropic news, October 15, 2026\.
* Anthropic Institute (2026). _Measurements for understanding the pace of AI development inside frontier labs_. Anthropic.
* Epoch AI (2026). _Epoch Capabilities Index_. Epoch AI.
* Anthropic, PBC (2026). _Form S-1/A Registration Statement_. U.S. Securities and Exchange Commission, filed November 2, 2026\.