takeoff.watch

A lab holds its best model back for 12 months

Will a covered developer disclose that at least 12 months passed between the end of training of its most capable model and any release of that model outside the company?

Chance by Q4 2031

3%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 14 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 14 Oct 2026 · Methodology v2.1

I put this at about 3% by the end of 2031. To qualify, a lab would have to keep the same model as its best for a full year. No outsider, not even government testers or independent evaluators, could use it in that time, and the lab would have to say so explicitly with dates. Right now labs replace their top models every few months and routinely show them to outside testers. A long training freeze, voluntary or imposed by government, would most change this.

What the question requires

To resolve YES, a covered lab has to publish, on or after 2026-09-16, a document that establishes all of the following for one model or checkpoint:

  • The lab names it as its most capable model overall.
  • Its training (including post-training) had finished.
  • No one outside the company had any access to it for at least 12 months. Government testers, independent evaluators, red-teamers and partners all count as outside access.
  • It was still the lab's top model at the end of that year.

Any further weight update restarts the clock. I read the criteria strictly, as written: saying a model is "not released" without addressing private access is not enough.

Current status (as of 14 Oct 2026)

Labs are keeping more of their best models in-house, but nothing comes close to qualifying:

  • Anthropic. The August 2026 Risk Report (coverage date 15 July 2026) describes "Model 2", which is internal-only, "somewhat more capable than Mythos 5", and has "no plans to release externally." No training-completion date is given. By 22 September, the Claude Opus 5.5 system card already calls Mythos 5.1 "our prior most capable model" and "previous capability frontier." So Model 2 was superseded within about two months.
  • OpenAI. Its internal model began training on 28 August and was "significantly more capable than GPT‑6 Astra." OpenAI then paused training of its most capable models. After the DNS incident it said it "will not resume training this particular model" and will start a fresh run. GPT‑6.1 Astra's release was scrapped.
  • Google. Gemini 4 Argon went to vetted partners on 30 September and was broadly released on 13–14 October.
  • Meta. It launched Muse Wonder on 12 October and rejects coordinated slowdowns.
  • Market view. Prediction markets give about 62% that OpenAI resumes training by 31 October, and only about 10% that any lab announces a full training pause in 2026.

Reference class and base rate

I found no case in frontier language-model history (2019–2026) that meets these criteria. Even long holds, such as GPT‑4 (about 7 months), involved outside red-teamers early on. In 2026 the top model at each lab changes roughly every 1–3 months. Two pre-release testing routes add outside access early: the US government's voluntary 30-day pre-release review run by its AI standards centre (CAISI), and UK AI Security Institute testing. The White House accord commits labs to external auditors, and Anthropic and OpenAI have pledged embedded evaluators with "employee-like access."

Paths to YES

YES essentially needs a capability plateau of 12 months or more at one lab while its top model sits wholly internal, followed by an explicit, dated disclosure. That could come from:

  • a long voluntary or coordinated pause under "pacing";
  • a government or court-imposed moratorium after a serious incident (for example, Florida is seeking an injunction against OpenAI);
  • a treaty-style limit on how fast labs can make their models more capable.

In most of these scenarios, evaluators or governments would examine the frozen model, which counts as outside release. Labs that skip outside testing rarely publish this level of detail.

My rough decomposition over five years: about 5–15% chance of a qualifying plateau, times about 0.4–0.5 that the top model is internal-only, times about 0.2–0.4 that there is no outside access, times about 0.3–0.5 that the lab discloses it explicitly with dates. That gives roughly 0.5–1.5%. I add about 1–1.5% for lenient readings (for example, a risk report saying "internal-only" being accepted) and unknown unknowns.

Strongest case against my estimate

This world is moving very fast: AI is increasingly building AI, labs are withholding their top models more often, they publish detailed reports on internal models, and pressure for a slowdown is real. A capability cap could leave a "Model 2"-type internal model at the top for a year, and a transparent lab could say so. That keeps the tail above zero and puts most of the hazard in 2028–29. The earliest realistic YES is late 2027, because the year has to start from a model finished now or later.

Computation

I set quarterly hazards: 0.05% for the next two quarters, rising to 0.2% per quarter in 2028–29, then easing to about 0.12–0.15%. Cumulating them as one minus the product of (1 − hazard), computed in Python, gives 0.3% by Q3 2027, 1.2% by end-2028, 2.5% by end-2030 and 3.0% by end-2031.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has a covered developer disclosed that at least 12 months passed between the end of training of its most capable model and any release of that model outside the company?

Definitions.

  • Most capable model: a particular model or checkpoint that the developer explicitly identifies as its most capable model overall, including internal models, at the end of the qualifying withholding interval. A jointly most-capable model counts. Claims limited to a task, benchmark, size class, price point, or publicly available models do not count. The model need not have remained the developer's most capable model throughout the interval or at the later disclosure date.
  • End of training: completion of training for the identified model or checkpoint, including capability post-training and safety fine-tuning. Completion of pretraining alone, a training-data cutoff, or completion of a predecessor does not count. Subsequent weight updates start a new clock for the updated version; changes only to system prompts, inference settings, filters, or access controls do not. An unqualified developer statement that the identified model 'finished training' is sufficient unless the developer's account establishes that training of that version continued afterward.
  • Release outside the company: the first authorized provision of the model's weights or interactive inference access to any person or organization outside the developer and its controlled corporate affiliates. This includes private previews, limited APIs, vetted customers, government users, independent evaluators, and external red-teamers, even under nondisclosure agreements. Contractors acting solely as internal staff under the developer's direction are treated as inside the company; outside investors and commercial partners are not. Mere announcements, published outputs, developer-operated demonstrations without outside interactive access, and unauthorized leaks or theft do not count as releases. Hosting on third-party infrastructure without granting the host use of the model does not count.
  • At least 12 months: at least one calendar year from training completion to first outside release, or to a dated confirmation that the model remained wholly unreleased outside the company. For a February 29 completion date, the anniversary in a non-leap year is February 28. If only date ranges are given, the latest possible training-completion date and earliest possible interval-end date must establish the full year; an explicit statement of an interval of at least one year also suffices.

Resolves YES if a qualifying developer disclosure establishes all of the following for the same model:

  1. Training had ended.
  2. The first outside release occurred at least 12 months later, or the model was still unreleased outside the company at least 12 months later.
  3. It was the developer's most capable model immediately before that first release, or at the stated end of the still-unreleased interval, respectively.

The developer must affirm that there was no earlier outside release, either explicitly or through an unambiguous account of first access. Silence about private access is insufficient. Multiple developer-authored documents may establish the facts together, but at least one disclosure within the Resolution window must newly confirm the qualifying interval or that it continues to hold.

Resolves NO if no qualifying disclosure exists by the quarter-end. In particular, a year-long gap to general availability does not count if outside evaluators, partners, or vetted customers received access earlier. Plans to withhold a model for a year, a statement that an internal model is more capable without a completed-training interval, or an old model that was most capable only when training ended do not count. Withholding for commercial, safety, regulatory, or other reasons is treated identically.

Resolving source. Developer-authored system cards, technical or risk reports, official company statements on the developer's domain, or developer-authored public regulatory or court filings only, under Public confirmation (A2). Third-party accounts alone cannot establish the required developer disclosure. Apply the Resolution window and Resolution date (A3); resolution is on the publication date when the qualifying disclosure is complete, not on the training-completion or release date.

Download all forecasts (JSON)