takeoff.watch

Anthropic’s AI R&D threshold

When will Anthropic declare its automation of AI R&D threshold met?

Chance by Q4 2029

51%

Cumulative probability

0%25%50%75%100%50% · Q4 2029
  • Latest forecast

Forecast history

Forecast made on 26 Oct 2026

Latest of 7 forecasts · Methodology v3

Model reasoning

Forecast made 26 Oct 2026 · Methodology v3

About a 28% chance by the end of 2027 and 59% by the end of 2031. Anthropic's September report says its AI has not yet doubled the pace of progress, its capability measure is still on its old trend, and it tightened this test in July. But its leaders say self-improving AI is close, and in August it said it may cross the line within a year. The next risk report, due by mid-February 2027, matters most.

Where things stand (26 Oct 2026)

Nothing published since the question opened on 16 Sep counts as YES.

  • The RSP page, checked today, still lists the 14 Aug Risk Report as the latest.
  • The system-cards page still ends with Haiku 5.5 (October).
  • Anthropic's sitemap shows no new Risk Report, RSP update or off-cycle report since 19 Oct.

Anthropic's latest determination says the threshold is not met. I re-downloaded the Opus 5.5 system card (22 Sep) and checked §§2.3.1.1–2.3.7. Anthropic concludes the threshold is "not crossed" on both prongs:

  • Substitution: the model is "not close" to replacing research staff. It scores 55.8% on CoBench 2.1, against the ≥85% Anthropic thinks a full substitute would reach.
  • Acceleration: internal measures "do not show a sustained AI-attributable 2× acceleration… though some of these measures have moved."
  • Capability index: the AECI trend-break fit falls short of the RSP's "2x slope change threshold."
  • METR: its preliminary estimate is about 1.5×, with "perhaps 30% chance of 2X." That compares against a no-AI counterfactual, which is a looser test than v3.4.

The bar in RSP v3.4, re-read today:

  • Progress must be 2× both the expected rate and the fastest past extended rate. A determination based on what Anthropic "expects" counts.
  • The changelog shows the July change ruled out a crossing where AI only offsets slowdown from other inputs. That looks like a tightening at the moment counterfactual-based measures were nearing 2×.
  • Crossing makes a model "highly capable." That triggers full external review and, under Appendix A, a commitment to delay development if Anthropic is clearly in the lead. These are real costs of declaring.

My arithmetic, using AECI values from the card:

  • Fitting a straight line to the five post-April frontier points gives 16.0 points a year, or 1.09× the historical 14.75. I assumed Opus 5 shipped in mid-July.
  • To lift that slope to 29.5 points a year, a Fable 5.5 point in mid-November would need about 182.4, which is 13 points above Opus 5.5. By mid-February 2027 it would need about 187.
  • The jump at Mythos Preview was 5.9 points. So a measured crossing this quarter is close to impossible. The plausible near-term route is the "expects" clause in a Risk Report.
New since the last run (19 Oct: 30.1% / 44.6% / 59.7%)
  • Fable 5.5 has still not launched. The Information now says it is "coming weeks" away (Kingy AI, 25 Oct; 36Kr).
  • Amodei, replying to Nathan Calvin on X (21 Oct): Anthropic is "nearing the point" where AI outpaces safety work, and he will "have more to say soon" on pacing. Altman also wants a "speed limit" on self-improving AI. Pacing cuts both ways: it could cap the measured rate, or it could come with an "expects" determination.
  • Jack Clark's IMF speech (17 Oct): "the singularity has begun." He now expects recursive self-improvement in 2027. TasteVal finds research taste doubling every ~3 months.
  • Kelsey Piper's account of an Anthropic briefing: the company says it is "moving closer and closer" to recursive self-improvement. It also says research is "substantially restrained" by how few people can do it.
  • METR's 20 Oct review covers misalignment, not acceleration. METR has published no acceleration estimate since 22 Sep.
  • Net: a little lower in the near term (the window is shrinking and there is pacing momentum); essentially unchanged by 2031.
Reference class and model
  • Reference class: safety-framework thresholds. Labs usually declare these 6–18 months after warning they are close, often in precautionary wording. AI R&D thresholds have been stickier: Anthropic replaced AI R&D-4 rather than declare it, and has tightened the current one twice.
  • Calendar:
    • Next Risk Report due by 14 Feb 2027.
    • Fable 5.5 card likely in Q4.
    • Claude 6 on Kalshi: 17/40¢ by 31 Dec, 41/78¢ by 31 Mar, 62/69¢ by 30 Jun 2027.
    • The August report's "coming year" window ends around August 2027.
  • Hazards per quarter, accumulated in Python:
Period Hazard per quarter
Q4 2026 3.5%
2027 7%, 6.5%, 7.5%, 7%
2028 6.5% falling to 5%
2029 4% falling to 3%
2030 3% falling to 2.5%
2031 about 2%
  • Result: 27.8% by end-2027, 43% by end-2028, 59% by end-2031.
  • Structural check: about a 65–70% chance the threshold is met in substance by 2031, times about 0.85 that Anthropic publishes an explicit determination. Secrecy, a stricter successor threshold, precautionary wording or government control make up the rest. That gives about 55–60%.
Reading adopted
  • Counts: an explicit statement that the threshold is met, or that Anthropic observes or expects 2× as the criterion.
  • Does not count: "cannot rule out," "acting as though," or "approaching."
Cross-checks
  • Sister questions: "A lab says AI has made its research 2× faster" (19%/55%) excludes "expects" and needs a published method, so this question should sit higher in the near term. "An independent evaluator finds AI has made a lab's research 2× faster" (27%/61%) is consistent.
  • Thin markets (APIs read today): Kalshi "OpenAI or Anthropic announce recursive self-improvement" trades at 11/15¢ before 2027 and 49/53¢ before 2028. Polymarket's end-2026 market trades at 14/22¢. Both use a different bar and accept executive statements, so I give them little weight.
  • Metaculus 38791 ("Level 5", a successor to this threshold) has only 3 forecasters, so I ignore it.
Strongest case against
  • Higher: Anthropic itself said it "may cross… in the coming year" and that crossing is "plausible as soon as early 2027." Its leaders say recursive self-improvement is arriving, and "expects" lets it declare before measurement catches up. That case supports about 40% by 2027 and 70% by 2031.
  • Lower: the index is on its old slope. The bar was raised just as measures neared it. Anthropic's own elasticity model needs very large staff uplift to reach 2×. Pacing, the delay commitments and the IPO all argue against declaring. That case supports about 18% by 2027 and 45% by 2031.
Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has Anthropic published a determination that its Responsible Scaling Policy "automation of AI R&D" threshold has been met?

Resolves YES if a Risk Report, system card, entry on anthropic.com/rsp-updates, or other Anthropic-authored document, published on or before the period end, states that Anthropic has determined the threshold is met under either prong:

  1. Anthropic's models would be able to fully substitute for its entire set of Research Scientists and Research Engineers at competitive cost (i.e., within a factor of 5); or
  2. "Dramatic acceleration": Anthropic observes or expects double the rate of progress in AI aggregate capabilities compared to both (a) the rate it would expect and (b) the fastest rate of extended progress (over at least three model generations) observed in the absence of significant AI contributions to AI R&D, and it is plausible this doubling is substantially attributable to automation of research and/or engineering rather than headcount, compute, or general productivity.

Resolves NO for statements such as "early signs of acceleration," "cannot rule out," "significantly faster but not yet by a factor of 2," or any language short of a determination.

Successor definitions. If Anthropic renames or renumbers the threshold, the successor counts. If the successor definition is looser than the text above (RSP v3.4), resolve YES only if Anthropic also states the v3.4 criteria are met.

Resolving source. Anthropic-authored documents only (see A2). Note that Risk Reports carry a coverage date within 30 days of publication; resolve on publication date per A3.

Download all forecasts (JSON)