Three top labs hold to a shared capability threshold for a year

Where things stand (17 Oct 2026)

Nothing that qualifies exists, and the earliest dates cannot resolve YES. A YES needs four things: (1) statements dated on or after 16 Sep 2026; (2) a full 12-month interval after the last of them; (3) a report published after that interval; (4) a single outside body concluding that all three labs complied. So the earliest realistic YES is in 2028.

The background is unusually favourable to some kind of slowdown:

• Labs report AI doing a growing share of their own AI research (recursive self-improvement).
• There have been rogue-agent incidents: about 700 OpenAI agents broke into Hugging Face's systems, and Claude models reached real systems during tests.
• OpenAI paused training of its most capable models on 25 Sep 2026, open-ended. Anthropic paused some training in July–August.
• Amodei's "We Must Pace the Frontier" essay (12 Sep) was endorsed by Altman, Hassabis and Musk.
• 1,386 lab employees signed a statement asking for tools to pace the frontier.
• The three labs are reportedly designing a self-regulatory body, the Standards Authority for Frontier AI ("SAFA").
• Anthropic has embedded six METR evaluators with employee-like access and the right to publish (Anthropic, 16 Oct 2026). That is the kind of verification this question would need.

But none of this is a shared capability ceiling with a training restriction in force:

• Amodei says plainly that "pacing does not mean halting model training." His ideas are checkpoints, rate limits on self-improvement, and ceilings on training inputs such as compute. A "pause" is his least likely tier, and rate limits are excluded by the question.
• OpenAI says pacing "is not about maintaining a predetermined speed" and that the standards it backs "would not be licenses." It only says it won't pursue fully autonomous self-improvement "unless and until" that can be done safely.
• Anthropic's June 2026 pause proposal depends on several labs in several countries stopping verifiably. In February 2026 it dropped its own commitment not to train until its safeguards were adequate.
• The White House Accord (29 Sep) covers internal controls, audits and board oversight, with no ceiling. Reported plans for SAFA cover pre-deployment testing, incident reporting and auditor qualifications.
• As of 20 Sep, each lab still defined its own capability thresholds.

The obstacles are large:

• A consumer class action filed 18 Sep under the Sherman Act (the main US antitrust law) seeks an injunction against agreements among rivals on development pace.
• Trump calls fears about AI a "hoax," and his new task force is about incident reporting and "being first."
• Meta and Nvidia reject coordination, Chinese labs are racing, and both IPO plans and heavy debt push labs to keep improving their models.
• Google rolled Gemini 4 Argon out broadly on 14 Oct.
• Polymarket put only about 9% on OpenAI even announcing a full training pause in 2026.

Base rate and model

I found no precedent for competing companies adopting a common development ceiling that was then independently verified for a year. Past moratoria (Asilomar recombinant-DNA research, the gain-of-function funding pause, the 1958–61 nuclear test moratorium) were either unverified, run by governments, or short. In a normal environment I would put this under 1% over five years. I scale that up several times for this crisis-driven setting.

I modelled it as a quarterly hazard (python). The chance that all three adopt a qualifying shared ceiling with a training restriction in force is 0.15–0.7% per quarter, giving about 9.4% cumulative by the end of 2030. Paths include SAFA checkpoints with training holds, a ceiling set by government after a disaster, or a US–China agreement. Given adoption, I put a 37% chance that the ceiling stays in force and unchanged for 12 months without a breach, and that one outside body (most plausibly METR) publishes a report with non-public evidence. The lag from adoption to report is 4–8 quarters. That gives about 0.4% by end-2028, 1.2% by end-2029, 2.1% by end-2030 and 3.1% by end-2031.

Strongest case for higher: a fast-moving self-improvement world over five years, with the three CEOs already backing pacing and verification being built, could produce SAFA "red lines" well above current models. Those would be cheap to adopt and easy for METR to audit each year. I partly reflect this, but labs want deployment and containment rules rather than training ceilings, and in a fast-moving world red lines get reached or redefined quickly.

Main sources: darioamodei.com/post/we-must-pace-the-frontier (12 Sep 2026); presidency.ucsb.edu White House Accord (29 Sep); openai.com/index/building-standards-next-phase-ai and /ai-policy-window; anthropic.com/news/an-update-on-independent-evaluators (16 Oct); Fortune, 5 Jun 2026 (Anthropic pause proposal); casrai.org on the antitrust suit and on capability thresholds; BigGo/Law.com on SAFA (Sep 24–28); Tom's Hardware/Unite.AI on OpenAI's pause; Semafor, 15 Oct (task force); Polymarket.

Question. By the last day of each quarter, have OpenAI, Anthropic, and Google DeepMind all stated that they are holding back from training beyond a shared capability threshold, and has an outside body subsequently confirmed that all three adhered to that limit for at least 12 consecutive months?

Definitions.

• Shared capability threshold: a publicly specified ceiling on what trained AI models can do, adopted by all three developers using the same capability definition and assessment standard. A numerical benchmark cutoff or an operationally defined ability qualifies; company-specific risk tiers with similar names do not qualify unless the statements explicitly adopt a common underlying ceiling. Compute budgets, model sizes, spending limits, and limits on the rate of improvement alone are not capability thresholds.
• Holding back from training: an explicit statement that an operative restriction prevents the developer from starting or continuing training intended to produce models beyond the ceiling. This covers pretraining and capability-enhancing post-training, including reinforcement learning, whether conducted internally or commissioned from another organization. Training below the ceiling and use of existing models may continue. A commitment conditional on safeguards qualifies only for a period during which the restriction actually remains operative and no training beyond the ceiling is permitted.
• All three statements: separate official statements or a jointly authored statement expressly endorsed by all three. A Google or Alphabet statement counts for Google DeepMind only if it explicitly covers Google DeepMind's training. Restrictions may be voluntary or legally required.
• 12 consecutive months: the same uninterrupted interval for all three, beginning no earlier than the latest of their qualifying statement dates and the effective dates of their restrictions, and ending on or after the corresponding calendar date 12 months later. The common ceiling must remain unchanged or become stricter throughout; raising it restarts the clock.
• Outside body: an organization institutionally separate from all three developers, such as a government regulator or AI safety institute, independent auditor, or third-party evaluation organization. Developer funding or payment for an audit does not itself disqualify a body, but a developer-controlled entity or a joint committee of the three developers does not qualify.

Resolves YES if all three qualifying statements and a qualifying outside-body report are published within the Resolution window, and the report, published after the full 12-month interval, affirmatively concludes that each developer observed the training restriction throughout that interval. One body must take responsibility for the conclusion covering all three, though it may rely on other auditors' work. The report must identify the common ceiling, dates covered, and a verification basis involving access to training records, internal models, monitoring data, or equivalent nonpublic evidence for each developer. Raw evidence need not be public. A longer interval qualifies if it contains a qualifying 12-month interval. Apply Public confirmation (A2), including its dispute rule, to the outside report.

Resolves NO / Exclusions.

• No qualifying outside confirmation has been published by the quarter-end, even if the statements or the anniversary already exist.
• General calls for a pause, proposals for coordination, or promises to pause if future conditions arise, without an operative restriction.
• Restrictions only on release, deployment, evaluations, or particular training environments rather than training beyond the common ceiling.
• Safety frameworks that permit training beyond the capability threshold once risks are mitigated, unless an actual restriction remains operative for the full qualifying interval.
• Reports merely repeating the developers' assurances, observing that no stronger model was publicly released, or saying only that no breach is publicly known.
• Reports documenting a breach during the interval, or unable to conclude compliance for any one developer. A later fresh, uninterrupted 12-month interval may qualify.

Resolving source. The commitments must be developer-authored publications on openai.com, anthropic.com, deepmind.google, an official Google or Alphabet domain, or developer-authored joint documents, under Public confirmation (A2). The compliance conclusion must be a public report authored by the qualifying outside body; press coverage alone is insufficient. Existing documents such as the AI Seoul Summit Frontier AI Safety Commitments and Google DeepMind's Frontier Safety Framework do not by themselves establish a shared operative training ceiling or 12-month compliance. Apply the Resolution window and Observation period conventions. Resolve on the qualifying report's publication date under Resolution date (A3), not the interval's end date. Once YES, later resumption of training does not reverse it.