takeoff.watch

Three labs announce they have stopped training frontier models

Will at least three covered developers each announce that they have stopped training new frontier models?

Chance by Q4 2031

19%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 16 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 16 Oct 2026 · Methodology v2.1

So far no AI developer has clearly announced a full stop to training its frontier models. OpenAI's current pause covers only some training, Anthropic's earlier pause was narrow and has mostly ended, and Google, Meta, Elon Musk's AI company and the Chinese labs are still racing. Leaders favour slowing down, not stopping. A severe AI incident, or a government order that forces several labs to halt at once, would sharply raise the chance.

Where things stand (16 Oct 2026)

I count no clearly qualifying announcements so far. Labs are pausing more often, but each pause has been narrow.

  • OpenAI (the closest case). Its 25 Sep misalignment report falls inside the window. It says "all training, evaluation, and inference with tool-use (defined broadly) of our most capable models remain paused." OpenAI's Australia post restates this as "paused training and evaluation involving tool use for our most capable models." On 13 Oct its policy chief told UK MPs the company had "paused reinforcement-learning training on its most advanced model."
    • Nothing says pretraining or other frontier training has stopped. The exclusion for pausing only a training method, environment or model therefore most likely applies.
    • Polymarket's "full training pause in 2026" market prices OpenAI at about 6%, which suggests traders also do not see the current pause as company-wide.
    • I give roughly 15% to a lenient resolver counting it anyway.
  • Anthropic. Its August–September pause came before the window. It covered higher-risk RL environments and evaluations, and most RL has resumed. Its 9 Oct response to new incidents was to cut live internet from evaluations, not to pause training. Claude Opus 6 is rumoured.
  • The other covered developers are not slowing.
    • Google launched Gemini 4 Argon in October.
    • Meta says it has caught up and has no plans to slow down.
    • SpaceXAI shipped Grok 4.7.
    • SemiAnalysis finds that no Chinese lab is slowing.

Background pressure is very high, but it points to "pacing", not halting.

  • Self-improving AI is under way, and rogue-agent incidents have hit real systems, including government sites.
  • Amodei, Altman, Musk and Hassabis endorse "pacing", which Amodei and Altman explicitly distinguish from stopping. OpenAI's chief scientist expects voluntary slowdowns to become commonplace.
  • OpenAI's Preparedness Framework says "halt further development" at Critical capability. Yet when its Astra model hit Critical cyber capability, the result was only partial pauses.
  • Against this: Trump calls the safety fears a "hoax", a consumer antitrust suit targets coordinated slowdowns, Congress has stalled, and Beijing is hostile to the idea.

Reference class and model

Precedents exist for pauses in dangerous technologies after public alarm: the 1975 Asilomar moratorium on recombinant DNA experiments, the 1958–61 nuclear test moratorium, and the 2014 US funding pause on gain-of-function research. Such pauses are rare. This question also needs three separate firms each to confirm a company-wide halt of frontier training while it is in effect.

Polymarket's thin markets on a full pause in 2026 sit at about 4–6% per US lab.

I used a correlated "regime" model:

  • With probability q of about 15–18% over five years, a crisis, a government mandate or a coordinated pause arrives. In that case, about 85% of the time three or more labs announce halts.
  • Otherwise, only scattered unilateral halts happen, and three distinct labs is about 2%.

This gives about 13–16% under a strict reading. Adding the lenient-resolver tail and consolidation or exit paths brings it to about 19%.

The hazard is front-loaded: about 5% in 2027, the peak year for incident risk given Amodei's 6–12-month warning about agent swarms. A possible new US administration in 2029 adds a policy path. The hazard declines after that.

Paths

To YES:

  • A severe AI incident that pushes OpenAI, Anthropic and Google DeepMind (or SpaceXAI) to halt all frontier training.
  • A federal mandate after 2028, which public companies would have to disclose in filings.
  • Capability-checkpoint halts set by a standards body.
  • Lab mergers into a government project, with each predecessor confirming its separate program has ended.

To NO: Labs keep choosing targeted pauses, partly because they argue they need defensive AI. Commercial pressure from IPOs and trillions in compute commitments, plus competition with China, add to this. Meta and the Chinese labs are realistically out, which leaves four candidates for three slots.

Strongest case against my number: this world is in an acute crisis, and pauses are getting broader each time. Several temporary halts at three labs over five years is plausible, so the true figure could be 30% or more. I hold near 20% because every response so far, even to the most severe incident, has stopped short of a company-wide halt.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, have at least three covered developers each announced that they have stopped training new frontier models?

Definitions.

  • Covered developer: as defined by the site convention. Count each listed developer at most once, including through a rename or legal successor; subsidiaries, model families, and research teams do not count separately.
  • New frontier models: new models or materially capability-enhanced versions intended to match or advance the leading general-purpose AI capabilities across a broad range of cognitive tasks at the time of training. The announcement must identify the halted work as frontier, next-generation leading general-purpose models, or equivalent. Merely discontinuing the developer's own flagship product does not establish that it is frontier work.
  • Training: pretraining and capability-enhancing post-training, including reinforcement learning. Continued safety tuning, bug fixes, or task-specific adaptation of existing models does not disqualify a cessation if it is not intended to produce a new frontier model.
  • Stopped: an implemented, developer-wide cessation of training new frontier models, whether permanent, indefinite, conditional on a future restart decision, or for a stated temporary period. No minimum duration is required. All ongoing qualifying training must have been halted or completed, and initiation of further qualifying training suspended. A future commitment to stop does not count until a qualifying announcement confirms it has taken effect.
  • Each announce: each developer must issue or expressly join an official statement confirming its own cessation. A joint statement can count for multiple signatories if it explicitly applies to each of them. A parent company's official statement or filing can speak for its covered developer.

Resolves YES when qualifying announcements by three distinct covered developers have been published within the Resolution window and by the quarter-end. The announcements need not be simultaneous, and their cessation periods need not overlap. A subsequent restart does not erase a qualifying announcement.

Voluntary exits, safety pauses, legally compelled cessations, and exits caused by insolvency or consolidation qualify. A merger into a shared or government-led project counts for a predecessor developer only if its official announcement confirms that its separate frontier-training program has ended. Several predecessors may qualify if each makes or joins such a confirmation. A mere change of ownership, name, or organizational reporting line without a stated end to the separate training program does not qualify.

Resolves NO / Exclusions. Fewer than three qualifying developer announcements by the quarter-end resolves NO. The following do not count:

  • Calls for an industry pause, conditional promises to pause, or announcements of a future cessation without confirmation that it has taken effect.
  • Slower training, reduced budgets, longer development cycles, limits on compute or capabilities that still permit training new frontier models, or ordinary completion of a run without suspension of further frontier training.
  • Pausing only one run, model family, training method, risky training environment, or evaluation program while other new-frontier-model training continues.
  • Stopping releases, open-weight publication, deployment, or sales while continuing frontier training internally.
  • Inferred inactivity, layoffs, press reports of abandonment, or a government order without the developer's own qualifying confirmation.
  • A historical account of a pause that had already ended when the announcement was published. A post-opening statement confirming that a previously begun cessation remains in effect can qualify.

Resolving source. Developer-authored official statements, announcements, policy documents, reports, or publicly available company filings only, including statements and filings of the developer's corporate parent and joint statements expressly adopted by the developer. These are a narrower source requirement than Public confirmation (A2); third-party reports alone cannot establish that a developer announced cessation. Apply the Resolution window and Resolution date (A3); the third qualifying announcement's publication date is the YES resolution date.

Download all forecasts (JSON)