Have AI timelines suddenly moved?

12 Nov 2026

A week ago METR put out a report on AI R&D acceleration at Anthropic.1 In the days that followed several follow-on analyses rapidly came out, including one by Ajeya Cotra,2 one from the AI Futures team,3 and a substantial post on X from Buck Shlegeris (CEO of Redwood Research).4

Together, they suggest that Anthropic has accelerated its capabilities progress since late 2025, and attribute the bulk of the improvement to increasing AI R&D capabilities (“AI building AI”). If true, this would be a big deal, since better models make a better research workforce, so this leads to compounding in capabilities.

The follow-on analyses all follow this compounding to bring forward their prediction for when AI will be superhuman at building AI, from 2028-2029 or beyond to 2027, next year.

The stakes are high, since this year’s OpenAI HuggingFace hacking incident5 demonstrates that we’re not ready for a massive capability jump, not in terms of how aligned models are and how well we can control them. So even an expert analysis is worth trying to understand and replicate.

Anthropic’s capabilities have started growing faster

Epoch AI provide the best reference data for capabilities improvements and the factors that go into them. They measure capabilities according to the Epoch Capabilities Index (ECI). Anthropic has their own internal version that includes private benchmarks, the AECI. On both of these measures, there’s a clear shift in the trend around October 2025.

Interestingly, this statistical fit for the hinge point arrives near the time of Opus 4.5. Many developers think about that release as a kind of tipping point in software: beforehand, most code needed to be written by humans; after, coding agents were good enough that using them for coding was the new default. To me it signalled the coming end of coding as a human craft.

The speedup in Anthropic’s own index, which includes both its own internal benchmarks, and unreleased models, is faster still, 1.9x on METR’s statistical fit, as low as 1.5x under other reasonable modelling choices. Since Mythos 5 in June 2026, every point that raised the internal frontier was an unreleased model, not available for the general public to evaluate.

How does this relate to what’s been going on in the other labs though? They also have strong coding models, are they experiencing a faster increase in capabilities than before?

The frontier isn’t growing faster than before

It turns out that competing labs have not sped up, if anything they may have slowed slightly. If you only take their best model releases, then the trend along the frontier has remained very stable, and this includes Anthropic’s models.

So, Anthropic has accelerated its capabilities growth, whilst other labs have not, but yet the frontier as a whole has not accelerated. Right now, Anthropic and the other labs are on par, leap-frogging each other for momentary lead, before the next model release from a rival steals the crown again.

We can see the longer run trend if we check the frontier model releases below:

We can see that post Oct 2025, we are still not exceeding the prior trend. Every new best model from any lab fell below the long term trend for most of 2026; only the recent Fable 5.2 release has landed back on trend. This doesn’t discount the possibility of a sustained Anthropic speed-up, it just shows that it is not yet certain.

What if Anthropic continued to accelerate?

To be clear, the existing trajectory of AI progress has been quite worrying. Today’s models are superhuman in some of their capabilities but well below human in others (“jagged”). But many predict we’d have superhuman AI researchers in the next 2-3 years under this prior trajectory.

Anthropic’s new pace of capabilities growth alone is not what worries researchers most; rather, it’s that the growth is AI driven, meaning that Anthropic might be entering a phase of continued, compounding acceleration, where the pace will keep rising. That brings estimates forward in their modelling, meaning 2-3 years becomes just 1-2 years.

Both of these timelines are worrying, but the second is a lot less time to develop the policy muscle to slow things down, and neither give enough time to put together the safety techniques and policies necessary to avoid catastrophic misuse or loss of control risks.

Will Anthropic continue to accelerate?

That largely depends on what’s caused the acceleration, and this is the most important question at the heart of each analysis of the situation.

  • ❌ It’s unlikely to be catch up growth, since other labs that were behind did not accelerate and catch up in this period.
  • ✅ Compute growth is a plausible factor, since Anthropic’s compute capacity grew about 5.7-fold in 2025, against 4.5-fold at OpenAI and 3.8-fold at Google DeepMind.6
  • ✅ AI building AI is a likely factor, since Anthropic reports that internal AI use grew 250x from Jan to Sep 2026,7 and reported a 12x speed-up of work by staff members in this period.1

So the entire question about Anthropic’s pace is what proportion of its growth is due to AI contributing more to building AI.

METR make a guess that AI building AI accounts for ~80% of the acceleration, using redacted information that outsiders can’t verify. Anthropic’s own public and private reports conflict regarding the extent of the speedup.89 Orgs outside Anthropic report 1.5-3x gains from AI, not 12x,10 and Anthropic’s pace didn’t apparently rise across 2026 despite 250x increased AI usage, so there are reasons to hold some skepticism here.

How will we know?

If Anthropic is not just keeping its new pace but accelerating it, we expect to see that reflected in the capabilities of their upcoming model releases.

Based on prior releases, we might guess that an Opus release might land in late November and a Fable release in late Jan. But at Anthropic’s current pace it wouldn’t overtake the prior frontier trend until Aug 2027. It’s only if Anthropic continues to accelerate that we’d expect to notice, probably not until the next Fable release. A release on time and several points above trend could be evidence of acceleration there.

We also expect to get a new METR report from a team without insider access to Anthropic, which reportedly found much less of a speed-up.

If the main source of Anthropic’s capabilities growth is truly AI building AI, it’s unclear why the other labs haven’t similarly benefitted. Other labs report heavy AI use too, with Google reporting 75% of new code being AI-generated.11 If other labs started accelerating their capabilities progress in a similar way, that could also provide some evidence.

Where does that leave us?

Even after re-running the modelling, I have trouble processing it and coming to a coherent view of what I believe here. I do believe the data to conclusively disambiguate these scenarios exists, but only within Anthropic.

  • If you believe that Anthropic’s capability growth is mainly compute related, then a superhuman researcher is most likely due in 2028 or 2029, the still shocking status quo scenario from top forecasters
  • If you believe that it’s mainly due to AI contributions to research, then the chance of a superhuman researcher in 2027 is much higher
    • ~70% if you accept METR’s headline figure, based on Anthropic’s internal evals of private models
    • ~40% if you accept only the public numbers by Epoch
    • ~20% if you accept only the public numbers and believe the appropriate reference point is the field’s earlier trajectory

Anthropic’s CEO has publicly stated the speed-up is primarily due to AI contributions to research.12 I find myself having trouble accepting that at face value, without substantially more evidence. Anthropic would serve the world better if it released enough information to conclusively show which scenario it is in, so that we can each take action appropriate to the moment.

Footnotes

  1. https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/ ↩ ↩2

  2. https://www.planned-obsolescence.org/p/ai-is-dramatically-accelerating-ai ↩

  3. https://blog.aifutures.org/p/q35-timelines-update-looks-like-the ↩

  4. https://x.com/bshlgrs/status/2120271165692461316?s=19 ↩

  5. https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/ ↩

  6. Epoch AI, AI chip owners dataset (accessed 11 November 2026), https://epoch.ai/data/ai_chip_users.zip. Anthropic’s fleet goes from about 209k H100-equivalents at end-2024 to 1.19M at end-2025. Epoch’s Anthropic 2025 figure is partly based on an estimate from an OpenAI investor memo. ↩

  7. Anthropic, PBC, Form S-1/A, Amendment No. 2 (filed 10 November 2026), p. 193: https://www.sec.gov/Archives/edgar/data/2130357/000162828026072484/anthropicpbc-sx1ax2.htm ↩

  8. Anthropic, Risk Report: August 2026 (redacted public version), pp. 11, 101–103. https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf ↩

  9. METR, “Review of the ‘Risks from automated R&D’ section in Anthropic’s Risk Report (August 2026)” (5 November 2026), the only public source for the private report’s wording: https://metr.org/blog/2026-11-05-rd-section-anthropic-risk-report-august-2026-review/ ↩

  10. Buck Shlegeris, “Where is all the AI safety research uplift?”, LessWrong (9 November 2026), puts Redwood Research’s speed-up at “more like 2-3x”: https://www.lesswrong.com/posts/wPrpnfaNL5JP9i6KT. Apollo Research’s reported ~1.5-2x is second-hand, via Zvi Mowshowitz. ↩

  11. Sundar Pichai, “Cloud Next ‘26: Momentum and innovation at Google scale”, Google blog (22 April 2026): https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/cloud-next-2026-sundar-pichai/ ↩

  12. https://darioamodei.com/post/we-must-pace-the-frontier ↩