An independent evaluator finds AI has made a lab's research 20× faster

Where things stand (October 2026)

There is already a way to publish such an estimate. The number it gives is far below 20.

• METR's current figure for Anthropic is about 1.5×. METR's summary of its Claude Opus 5.5 evaluation (22 Sep 2026, https://metr.org/blog/2026-09-22-claude-opus-5-5/) was produced under an AI R&D assessment agreement with Anthropic. It cites a separate METR team with "elevated access" whose preliminary estimate is "~1.5X overall acceleration in capabilities due to AI (i.e. 1.5 years in 1 year), with perhaps 30% chance of 2X." It says more public output is expected "in the coming weeks."
• That is exactly the kind of number this question asks about, so whoever would publish the resolving estimate already has access and is already publishing. The issue is magnitude.
• Anthropic's own view is also below 2×. Its August 2026 Risk Report (https://www.anthropic.com/aug-2026-risk-report) says acceleration is "less than a factor of 2" and below its 2× early-warning threshold.
• A METR staff note puts researcher uplift near 2.5×. Thomas Kwa (8 Jul 2026, https://metr.org/notes/2026-07-08-anthropic-researcher-uplift/) estimates about 2.5×. Under Greenblatt's production function that maps to roughly 1.8× overall R&D speed-up.
• Other signs of rapid automation, but not of a 20× pace:
  • Claude "leads" 26% of Anthropic's AI R&D (https://www.anthropic.com/institute/measuring-pace-of-ai-development, 17 Sep 2026).
  • OpenAI reports 3.1 agent-workdays per human workday and targets an automated AI researcher by March 2028 (https://openai.com/index/research-acceleration-view-inside-openai/).
• Independent access is expanding:
  • Anthropic has committed to embedded evaluators; Faculty/Accenture is first, and METR is in talks.
  • OpenAI says it is finalising contracts with outside assessors (CNBC, 11 Oct 2026).
  • Ryan Greenblatt has joined METR to study how close labs are to research automation feeding back into faster progress.

Reference class and base rates

20× means 20 years of progress per year. That is the "superhuman AI researcher" or intelligence-explosion regime, not incremental uplift. Views on when, or whether, it arrives:

• Experts and superforecasters (Aug 2025): in the METR/Forecasting Research Institute pilot (https://metr.org/blog/2025-08-20-forecasting-impacts-of-ai-acceleration/), experts gave 20% and superforecasters 8% to a mere 3× acceleration of effective-compute growth by 2029.
• AI Futures Project: its August 2026 model runs, which assume no deliberate slowdown, put the median for superhuman-researcher-level research skill (about 17–25× in its model) at roughly 2028.9 (Kokotajlo), 2029.5 (Brendan) and about 2030.8 (Lifland). Source: https://blog.aifutures.org/p/q25-2026-timelines-update-uplift and its forecast pages.
• Greenblatt puts full automation of AI research around November 2028.
• Sceptics: Noam Brown says about 3× is plausible and 10× unlikely. Ramez Naam and the "Economics of RSI" paper find the feedback loop is not yet strong enough to sustain itself. Anthropic's internal production-function view (2× labour gives 1.15–1.3× output) implies 20× needs enormous labour multipliers.

Model

I built the forecast in Python as a mixture of three views of when the true multiplier would reach 20× at a lab with independent access, absent any slowdown:

• aggressive (AI Futures/Greenblatt), weight about 0.33, cumulative 67% by end-2031;
• mainstream, weight 0.45, 18% by end-2031;
• sceptical, weight 0.22, 3% by end-2031.

The mixture gives about 0.31 by end-2031. I then applied:

• Pacing: 42% chance of no effect, 43% chance of a delay of about 1.5 years, 15% chance of a hard cap. Pacing has strong momentum at Anthropic and OpenAI, and OpenAI's framework halts development at about 5×. But the White House opposes slowing and China will not pace.
• Publication: a 0.62 chance that, once the true multiplier reaches 20×, an independent organisation publishes a central estimate of at least 20, with an exponential lag averaging about 4 months. This allows for conservative methods, ranges being read at their lower bound, capability-based measures that come out lower than research-only ones, and possible secrecy.
• False positives: about 1% spread over the window.

The result is about 0.5% by end-2027, 2% by end-2028, 5% by end-2029, 9% by end-2030 and 13% by end-2031. Hazards are near zero now and rise through 2029–2031.

Strongest case the other way

Insiders (Jack Clark, Kokotajlo, Greenblatt, OpenAI's Pachocki) expect research to automate itself around 2028. Once that happens, the multiplier could pass 20× within months, while embedded evaluators are in place to measure it. That case argues for 25–35%. I discount it because the best independent measurement is about 1.5× and labs' internal success horizon on research tasks is still short. Compute bottlenecks and pacing also point the same way. On the other side, superforecaster-style priors would argue for about 5%.

Question. By the last day of each quarter, has an independent organisation published an estimate, giving a number and a method, that AI systems have made the overall pace of a covered developer's AI research at least 20 times faster than it would be with human researchers alone?

Definitions.

• Independent organisation: a named organisation that is not part of, controlling, controlled by, or under common control with the covered developer concerned (control meaning a majority of equity, votes, or board appointments), and neither it nor any entity controlling it has made a direct equity investment in that developer. A government body is assessed apart from its government's other holdings. Paid work for the developer, compute or model access, grants, donations, and the developer's pre-publication review or redaction do not affect independence; an author employed by the developer does. It must have produced the estimate with access to the developer's non-public research data, staff interviews or surveys, or unreleased models under an agreement or memorandum of understanding with the developer or a legal or regulatory requirement, as stated in a publication by either.
• Overall pace of a covered developer's 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 organisation explicitly presents it as an estimate of the developer's 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 20 times faster: a ratio of AI-assisted research progress per unit time to the human-only counterfactual of at least 20.0, or equivalent time to the same progress of at most one-twentieth. The AI systems need not have been built by the developer whose research is estimated.
• 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 organisation 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 resolving source gives, as the independent organisation's own estimate, an achieved or currently observed overall multiplier of at least 20.0, explicitly applied to a named covered developer, with the method published there or in public documentation it cites. A central or point estimate counts despite an uncertainty interval extending below 20.0; “about,” “roughly,” or “~” figures are read at face value; a range given alone is read at its lower bound; and if several estimates for the same developer and period include none identified as central or overall, the lowest governs. Preliminary or experimental estimates count.

Exclusions (resolve NO).

• Multipliers for coding speed or volume, experiment counts, or agent labor (such as agent-workdays per human workday).
• Unadjusted averages or medians of staff's self-reported output gains.
• Figures such as “up to 20x” or “nearly 20x,” and a stated probability that the multiplier reaches 20.

Resolving source. Only the organisation's website, including staff-authored notes hosted there; reports, papers, preprints, testimony, or submissions under its name or affiliation; and sections of developer-published documents explicitly presented as its own text, not a developer's paraphrase, summary, or quotation. This replaces Public confirmation (A2): developer disagreement does not prevent YES, but the organisation withdrawing the estimate, or revising it below 20.0, within 14 days of publication does. Apply the Resolution window, Observation period, and Resolution date (A3) conventions; the date is when the estimate and its method are both public.