A lab says AI has made its research 2× faster

Where things stand (17 Oct 2026)

No covered developer has yet published a qualifying statement of 2× or more. The nearest statements all fall below 2×, are excluded proxies, or come from third parties:

• Anthropic, Risk Report August 2026 (published 14 Aug; data as of 15 Jul): "Our leading indicators point to … meaningful acceleration starting in early-to-mid 2025, though by less than a factor of 2… the overall pace of progress remains below the RSP threshold of a doubling over pre-AI-assisted rates." The leading indicators are redacted. The report also says "we may cross this threshold in the coming year" and that the threat model could become "a major concern in the next 6–12 months."
• Claude Opus 5.5 System Card (22 Sep): "our internal measures do not show a sustained AI-attributable 2× acceleration… though some of these measures have moved." Anthropic's capability index (its fork of Epoch's capability index) gives a slope change of 14.4→22.2 points/yr, which is 1.53× (95% range 1.20–1.82). A one-time jump fits the data better than a slope change. The card also quotes METR's preliminary estimate of "~1.5X overall acceleration in capabilities due to AI… perhaps 30% chance of 2X." That figure comes from a third party, gives no observation period, and was published before the question opened, so it does not count.
• Anthropic's Responsible Scaling Policy, version 3.4 (8 Jul) defines this threshold almost exactly as the question does: twice the rate of aggregate-capability progress compared with the pre-AI baseline, "substantially attributable" to automation rather than headcount or compute. So the measurement machinery already exists, and Anthropic reports on it in every system card and Risk Report.
• Excluded proxies: Claude "leads" 26% of Anthropic's R&D work (Anthropic Institute, 17 Sep); engineers merge 8× as much code per day (Q2 2026 vs 2024) and staff self-report about 4× output ("When AI builds itself"); OpenAI reports 3.1 agent-workdays per human workday and explicitly gives no overall pace figure ("Research acceleration: the view inside OpenAI," early Sep). OpenAI's GPT-6 system cards (Sep–Oct) say AI self-improvement is below its High threshold.
• Google DeepMind, Gemini 4 Argon safety report (Oct): "no significant end-to-end acceleration of AI R&D." Its capability-index slope went from 0.091 to 0.118 points/month, which is not statistically significant, and it attributes Argon's above-trend score to extra compute and investment.
• Leaders' personal statements, which do not count: Dario Amodei (12 Sep) says progress is "drastically faster… driven primarily by" AI building AI; Jack Clark (12 Oct) now favors 2027 for full recursive self-improvement.

Reference class and adjustments

Labs have usually declared safety-framework thresholds crossed within about 6–18 months of first saying they were close: Anthropic's ASL-3, OpenAI's High bio, High cyber (Feb 2026) and Critical cyber (Sep 2026). Anthropic now says this about the AI R&D threshold itself. Two things pull the probability down. No lab has ever published a quantitative AI-attributable multiplier of 2 or more. And the question excludes the proxies labs prefer to publish: code volume, agent-days, automation share and self-reports.

Paths to YES

1. An Anthropic Risk Report (next due between mid-Nov 2026 and mid-Feb 2027, then every 3–6 months) or a system card declares a sustained AI-attributable doubling, with index slopes and an attribution argument.
2. Anthropic or OpenAI publishes its own counterfactual estimate. Both are pushing standards for measuring progress toward recursive self-improvement, and METR now has employee-like access at Anthropic.
3. Google DeepMind finds a doubled slope that it attributes to AI.

Paths to NO

• Measured capability trends lag and may stay linear because of compute bottlenecks; DeepMind's flat trend is evidence for this.
• A declaration may be framed as "expect" or "cannot rule out" (both forecasts), or rest on redacted indicators (no method).
• Only partial attribution: a doubling that is just "substantially" AI-driven may leave the AI-attributable part below 2.
• Thresholds may be redefined; pacing agreements may cap speed; information may become secret; progress may stall.

How I read the question

I count an Anthropic statement that a sustained AI-attributable doubling has occurred if it comes with slope figures and explains why compute and headcount do not account for it. I do not count METR-written text inside Anthropic documents unless Anthropic adopts the number and method itself.

Forecast construction

I used quarterly hazards, computed in Python: 3.5% (Q4 2026); 5.5–7% through 2027, peaking after Anthropic's "coming year" window; 5–6.5% in 2028; 3.5–4.5% in 2029; 3% in 2030; 2.5% in 2031. Cumulative probability is about 26% by end-2027, 42% by end-2028, 51% by end-2029 and 61% by end-2031. As a check, scenario weights of rapid self-improvement 35%, steady progress 35%, stall 20%, pacing or secrecy 10% give about 60% by 2031.

Strongest case against: labs consistently avoid counterfactual multipliers, and DeepMind sees no acceleration. That would support roughly 35–40% by 2031. I stay higher because Anthropic's policy is built around exactly this 2× comparison, and Anthropic already publishes the ratio (1.53×, with a confidence interval).

Sources

• https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf (22 Sep 2026)
• https://www.anthropic.com/aug-2026-risk-report (14 Aug 2026)
• https://www-cdn.anthropic.com/files/4zrzovbb/website/0bacdc8440ea96e62a8766d99ebe1d4eea6d5f3a.pdf (Responsible Scaling Policy v3.4, 8 Jul 2026)
• https://metr.org/blog/2026-09-22-claude-opus-5-5/
• https://metr.org/blog/2026-10-14-ai-rd-acceleration-assessment-proposal/
• https://www.anthropic.com/institute/measuring-pace-of-ai-development (17 Sep 2026)
• https://www.anthropic.com/institute/recursive-self-improvement
• https://openai.com/index/research-acceleration-view-inside-openai/
• https://deploymentsafety.openai.com/gpt-6-astra
• https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf (Oct 2026)
• https://darioamodei.com/post/we-must-pace-the-frontier (12 Sep 2026)
• https://importai.substack.com/p/import-ai-476-creativity-and-rsi (12 Oct 2026)

Question. By the last day of each quarter, has a covered developer stated, giving a number and a method, that AI systems have made the overall pace of its own AI research at least 2 times faster than it would be with human researchers alone?

Definitions.

• Overall pace of its own AI research: the developer-wide rate of substantive progress in developing or improving AI models and methods, measured per unit of calendar time. For a covered developer that is a named AI lab or model organization within a larger company, the scope is that organization's AI research, not the parent's unrelated work. An aggregate estimate from a sample counts only if the developer explicitly presents it as an estimate of its overall AI research pace.
• Human researchers alone: a counterfactual in which the same human research workforce has ordinary non-AI software and research infrastructure but no AI systems performing or assisting research work. This does not exclude training or evaluating AI models as objects of research. The comparison must distinguish AI assistance from changes in human staffing, training compute, or other non-AI resources; a historical baseline counts only if the method explains how those differences are accounted for.
• At least 2 times faster: a ratio of AI-assisted research progress per unit time to the human-only counterfactual of at least 2.0, or equivalent time to the same progress of at most one-half. The AI systems need not have been built by the developer making the statement.
• Giving a number and a method: the publication reports the multiplier (or quantities explicitly equated to it), identifies the research-progress measure and observation period, and explains how both the AI-assisted result and human-only counterfactual were estimated and aggregated to the developer-wide level. Experimental comparisons, observational analyses, and explicitly described estimation models may count; public raw data, independent replication, and statistical significance are not required. Survey evidence counts only when the developer uses a disclosed method to translate it into and endorse an overall research-progress estimate, rather than merely reporting respondents' output estimates.

Resolves YES if a qualifying developer-authored publication states that an achieved or currently observed overall multiplier is at least 2.0 and supplies the method above, directly or through linked developer-authored documentation. A stated central estimate of at least 2.0 counts despite an uncertainty interval extending below 2.0. If only a range is given, its lower bound must be at least 2.0. The Observation period convention applies.

Exclusions (resolve NO).

• Forecasts, targets, hypothetical scenarios, claims of potential acceleration, and model capabilities not translated into achieved research acceleration.
• Multipliers for coding speed, code volume, experiment counts, agent labor (such as agent-workdays per human workday), benchmark task performance, individual researchers, selected teams, or selected projects without the required developer-wide research-progress estimate.
• An unadjusted average or median of employees' self-reported output gains; raw output is not itself overall research progress.
• Faster model releases, improved model scores, or a before-and-after increase in research pace without an explicit AI-attributed comparison against the human-only counterfactual.
• Statements such as “up to 2x,” or a numerical claim accompanied only by an assertion that internal analysis was performed, without the specified method.
• In the absence of a qualifying statement by a horizon, that horizon resolves NO.

Resolving source. Covered-developer-authored publications only: official website publications, research papers, system cards, risk reports, or public regulatory filings. Public confirmation (A2) applies with this narrower author requirement; third-party estimates do not independently qualify. Apply the Resolution window and Resolution date (A3) conventions. If the claim and required method are published separately, use the publication date on which both first become publicly available in qualifying sources.