Prolepsia forecastYes / no

On 31 December 2027, will the highest Epoch Capabilities Index score of any open-weight model be at least as high as the highest score of any model on 1 January 2027?

As of 17 November 2026, Prolepsia puts the chance of yes at 76%.

76%chance of yes

Yes 76%No 24%
Prolepsia forecast made

How Prolepsia reasons

The report Prolepsia wrote with this forecast.

How forecasts are made · Track record

TL;DR

The forecast is 76.3% Yes and 23.7% No. Open-weight models are favored to reach the beginning-of-2027 capability frontier by year-end; they do not have to match the best proprietary model available at the end of 2027.

Key evidence

  • Historical catch-up fits within the available year. Epoch’s May 29, 2026 analysis found a six-month lag when catch-up required an open model’s point estimate to exceed the earlier frontier. That is more relevant than its four-month headline, which used an uncertainty-based comparison (Epoch analysis).
  • The current hurdle is substantial but within recent progress. Rounded scores put Claude Fable 5.2 at 173 and the leading open-weight model, Kimi K3.1 Pro, at 161—a gap of about 12 ECI points (Epoch leaderboard). The January baseline is not fixed yet.
  • Open models have demonstrated gains of the required scale. Kimi’s displayed score rose from 145 for K2 Thinking to 161 for K3.1 Pro across approximately a year. Progress was uneven, so this is a pace check rather than a dependable annual growth rate (Epoch model data). Epoch’s broader reasoning-frontier trend of 16.6 points per year provides context, not an open-weight forecast (methodology report).
  • Release policy is a real constraint. Alibaba’s November 16 announcement offered hosted Qwen4-Max access while promising smaller open-weight models. A developer’s technical progress therefore need not appear in its strongest downloadable checkpoint (Alibaba announcement).

Uncertainties

The largest unknowns are the final January baseline, the pace of improvement in released weights, and relative score changes under the eventual common methodology. A major release before January raises the hurdle. Compute constraints, weaker capability transfer, or selective withholding could slow open-weight progress across several developers at once.

Historical comparisons favor Yes, but overlapping observation windows are not independent trials. The recent pace also includes both large jumps and quieter periods.

Resolution uses qualifying release dates and the official dataset available on 31 March 2028. Later evaluation can count a model released before the deadline; it cannot rescue weights first released afterward. Published point estimates—not overlapping confidence intervals—determine whether the target is met.

Read the full report

TL;DR

The forecast is 76.3% Yes and 23.7% No. The central expectation is that the best qualifying open-weight model reaches the 1 January 2027 frontier by 31 December 2027. The decisive distinction is that the target freezes in January, while progress counts only when sufficient weights become publicly downloadable.

Context

As of the November 17, 2026 forecast date, Epoch’s rounded scores put the overall leader, Claude Fable 5.2, at 173 and the leading open-weight model, Kimi K3.1 Pro, at 161. That leaves a gap of about 12 ECI points before any further increase in the January baseline (Epoch leaderboard). The question is about crossing a fixed capability threshold, not taking first place on the leaderboard at the end of 2027.

The comparison uses general ECI point estimates under one common methodology. The baseline includes models publicly usable by 1 January 2027; the year-end group requires qualifying weights released by 31 December 2027. Commercial restrictions do not automatically disqualify weights, but API-only access does. The official dataset available on 31 March 2028 supplies the scores, including retrospective evaluations. Exact equality counts as Yes.

Evidence

The strongest historical anchor is Epoch’s published open-versus-closed comparison. Its May 29, 2026 analysis covered January 1 through May 28 and reported an average gap of eight ECI points. The headline lag was four months, but the stricter comparison—requiring the open model’s point estimate to exceed the earlier frontier’s point estimate—gave six months. That is the relevant reference class for this question, because resolution does not accept statistical indistinguishability as equality (Epoch analysis).

A year gives more room than that historical lag. The historical annual comparisons also favor catch-up: open models exceeded the beginning-of-year frontier in 2024 and 2025, and had already exceeded the beginning-of-2026 frontier by the forecast date. These comparisons support the direction of the forecast, not a precise success frequency. They share models and periods of progress, and the available reconstructions differ in their treatment of dates and score vintages (Epoch historical score data).

The same caution applies to rolling windows. Many successful overlapping windows can reflect a few important releases rather than many independent tests. The useful conclusion is that catching a year-old frontier has recent precedent. It is not that the past supplies enough independent trials to make failure negligible.

A more concrete pace check comes from the Kimi family. The full selected sequence in the current-vintage observations is shown below. These are displayed, rounded general ECI scores. The dates identify model or API releases, not necessarily the dates when weights first became downloadable (Epoch score data; K2 Thinking model page).

Model Model/API release date Displayed general ECI
Kimi K2 Thinking 6 November 2025 145
Kimi K2.5 27 January 2026 149
Kimi K2.6 20 April 2026 152
Kimi K3 16 July 2026 159
Kimi K3.1 Pro 23 October 2026 161

This sequence shows a 16-point rise across approximately a year. It also shows why a straight-line extrapolation is fragile: only two displayed points separate K3 and K3.1 Pro. These are related releases from one developer, not a complete history of the open frontier. They establish that gains of the required scale have happened, while leaving substantial uncertainty about their timing.

Epoch’s November 4 methodology report gives a broader scale check. It estimates reasoning-frontier growth at 16.6 ECI points per year, with a reported 90% interval of 14.6–18.7. That is a historical trend for the overall frontier, not a forecast of downloadable-model progress (Epoch methodology report). The distinction matters: proprietary progress can move faster than released open-weight capability.

The current gap makes the task plausible without making it easy. As an illustration, adding an assumed two-point increase in the January baseline to today’s roughly 12-point gap requires about 12.5 points per year over the remaining time. The two-point increase is an assumption, not an observed release or a prediction of the final baseline. The required pace sits below the broader historical trend, but a slowdown or an unusually strong pre-January release could remove that room.

The most direct downside evidence concerns release choices. Alibaba’s November 16 Qwen4-Max announcement confirmed hosted access while promising smaller open-weight Qwen4 models. It did not establish a release of Max weights. This is evidence that a developer can keep advancing technically without advancing its strongest qualifying open-weight model at the same rate. It is not proof that Max will remain closed permanently (Alibaba announcement).

Safety policies create a similar distinction. Moonshot described limiting K3.1 Pro’s cyber capabilities before release and avoiding an open release that extends the absolute cyber frontier. Yet it still released K3.1 Pro. The evidence supports selective restraint, not a demonstrated ceiling on general ECI or a prohibition on matching year-old capabilities (Moonshot technical blog).

Capability transfer is another possible bottleneck. Anthropic reported that its defenses reduced harvested reasoning data by 90–99%. Its controlled example achieved 80.4% on SWE-bench Verified with full reasoning transcripts versus 71.8% with summaries. These are vendor-reported findings about a specific mechanism, not a measured reduction in general open-model progress. They nevertheless identify a reason historical catch-up could slow (Anthropic report).

Compute policy adds a potentially shared constraint. BIS announced restrictions addressing third-country remote access to advanced computing on November 16. The exact legal effective time and individual developers’ exposure were not established. The relevant risk is disruption across several developers, not a justified assumption that any particular model program must stop (BIS announcement).

There is counterevidence to a broad slowdown. DeepSeek’s October update improved benchmarks within an existing generation, while Z.ai attributed GLM-5.3 gains to post-training on the same base as GLM-5.2. These vendor results do not translate directly into ECI points, but they show that progress need not wait for a new large pretraining run (DeepSeek model card; GLM model card). Meta’s Muse Spark Open also supplies a route outside the leading Chinese developers, though its existence alone does not establish that it will reach the target (released model).

What's non-obvious

A widening live gap can coexist with success on this question. Proprietary models can pull further ahead during 2027 without changing the threshold once January has passed. Only one qualifying open-weight model needs to cross that frozen threshold. A previously proprietary model can also qualify if its weights become public during 2027. Neither route requires the whole open ecosystem to keep pace with the contemporary frontier.

The other subtlety cuts against an overly optimistic reading. The familiar four-month catch-up headline uses a looser test than resolution does; the strict historical comparison was six months (Epoch analysis). At the same time, restrictive licenses can still qualify under this question. The decisive checks are independent inference, actual weight availability, and an official general ECI score—not whether a release is described as fully open source.

Uncertainties

The January baseline remains unknown. A strong public release before the cutoff can raise it, and a qualifying model can receive its official score later. There is no verified schedule establishing the final maximum. The most useful new evidence would be confirmed public availability and official general ECI scores for late-year releases.

The next gap is the pace of released capability. Continued technical gains are insufficient if the strongest checkpoints remain hosted, depend on undisclosed components, or arrive as weights after the deadline. Binding release commitments and verified downloadable artifacts would clarify this risk. Rumored future models do not close it.

Scoring also remains material. Epoch adopted Bayesian general ECI estimation on November 4, 2026, with an inclusion requirement of at least four normalized benchmark scores between 0.1 and 0.9 (methodology changelog). A common methodology prevents artificial comparisons between incompatible scales, but additional benchmark evidence can still change relative point estimates. Today’s rounded scores are therefore neither the final comparison nor immutable floors.

Finally, the historical record is short and dependent. It supports catch-up within the available time, but does not establish how resilient that pattern is to simultaneous compute, transfer, and release-policy constraints. Those risks explain why Yes is favored without treating the outcome as assured.

Forecast history

How Prolepsia’s forecast has movedOne forecast so far: 76% on 17 Nov 2026.
DateForecast
76% chance of yes

Sources Prolepsia read

  1. Epoch Capabilities Index | Epoch AIepoch.ai
  2. Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity, and Discovery - Alibaba Cloud Communitywww.alibabacloud.com
  3. Kimi K3.1 Pro Tech Blog: Advancing Open Intelligencewww.kimi.ai
  4. Strengthening our defenses against distillation attacks \ Anthropicwww.anthropic.com
  5. BIS announcementwww.bis.gov
  6. deepseek-ai/DeepSeek-V4.1-Flash-1017 · Hugging Facehuggingface.co
  7. Kimi K3.1 Pro Tech Blog: Advancing Open Intelligencewww.kimi.com
  8. CAISI’s Assessment of Z.ai’s GLM-5.3 Cyber Capabilities | NISTwww.nist.gov
  9. Introducing Mistral Large 4 | Mistralmistral.ai
  10. Muse Wonder: A Frontier Model For Personal & Professional AI Agentsresearch.meta.ai

The question

This question asks whether, by 31 December 2027, the highest-performing open-weight AI model will have achieved an Epoch Capabilities Index (ECI) score equal to or greater than the highest ECI score achieved by any AI model available on 1 January 2027, regardless of whether that model was open-weight or proprietary. The question measures whether open-weight AI models can catch up to the general capabilities of the overall AI frontier at the beginning of 2027. The Epoch Capabilities Index, developed by Epoch AI, combines results from more than 50 AI benchmarks into a single measure of general AI capabilities. The index allows comparisons across different models and generations of AI systems, including open-weight and proprietary models. Epoch AI publishes these scores through its official ECI leaderboard at https://epoch.ai/eci. As of 17 November 2026, Epoch AI uses a Bayesian statistical methodology for calculating ECI scores, following a methodology update in November 2026. The index estimates model capabilities using performance across benchmarks of varying difficulty. Epoch AI also publishes comparisons between open-weight and proprietary models, making the index suitable for tracking the gap between these categories. The question compares two quantities: the highest general ECI score among all models publicly released on or before 1 January 2027, and the highest general ECI score among open-weight models whose weights have been publicly released on or before 31 December 2027. Models developed anywhere in the world are eligible. Models released before 2027 remain eligible for the end-of-year comparison, provided their weights are publicly available by the deadline. The comparison will use the same version of Epoch AI's general ECI methodology for both groups to avoid distortions arising from changes to the index over time. The question concerns model capabilities rather than whether the models occupy the same leaderboard positions on the respective dates.

How it resolves

The question resolves YES if the highest general Epoch Capabilities Index score among qualifying open-weight models available by 31 December 2027 is greater than or equal to the highest general ECI score among all qualifying models available by 1 January 2027. Otherwise, it resolves NO. The primary resolution source will be Epoch AI's official Epoch Capabilities Index leaderboard (https://epoch.ai/eci), supplemented where necessary by its downloadable datasets, documentation, and model release records. The baseline will be established by identifying every model publicly released on or before 1 January 2027 and determining the highest general ECI score among those models. Both proprietary and open-weight models qualify for this baseline. Publicly released means that the model was available for use by external users, either through an API, publicly accessible interface, or downloadable weights. Models that were only announced, privately tested, or demonstrated without public access do not qualify. The end-of-year comparison will identify every model whose weights were publicly downloadable by 31 December 2027 and determine the highest general ECI score among these models. An open-weight model is defined as a model for which the developer or an authorized distributor has publicly released sufficient trained weights to permit independent inference without requiring access to the original developer's API or infrastructure. Models with restrictions on commercial use or redistribution still qualify if their weights can be downloaded and independently executed. Models available exclusively through APIs, hosted interfaces, or inaccessible proprietary weights do not qualify. Both dates are inclusive, using 23:59:59 UTC as the cutoff for determining model availability. A model qualifies based on its release status by the relevant cutoff, even if Epoch AI publishes its ECI score afterward. To ensure comparability, resolution will use the general ECI scores calculated under a common methodology, rather than comparing scores captured from different historical versions of the leaderboard. The official Epoch AI dataset available on 31 March 2028 will be used, including scores for qualifying models added retrospectively before that date. If Epoch AI publishes a subsequent correction before resolution is finalized, the corrected scores will be used. Domain-specific ECI variants, including software engineering and mathematics, will not be considered. The comparison will use Epoch AI's published point estimates rather than confidence intervals. Scores will be compared at their maximum available published precision, without additional rounding. Exact equality qualifies as YES. The question resolves NO if at least one qualifying model has a baseline score, at least one qualifying open-weight model has an end-of-year score, and the highest end-of-year open-weight score is strictly lower than the baseline maximum. The question will be annulled if Epoch AI permanently discontinues the general ECI or changes its methodology in a way that prevents a meaningful comparison using a common scale, and no suitable official historical dataset or recalculated scores are available. It will also be annulled if the information necessary to establish either maximum remains unavailable by 31 March 2028. Missing scores for individual models do not automatically require annulment; the comparison will ordinarily use the qualifying models for which Epoch AI provides scores, but annulment is appropriate if missing coverage makes the relevant maximum indeterminable.

Fine print

The question concerns the capabilities of individual AI models, not the average capabilities of open-weight models or the cumulative capabilities of all models available in either period. A model released before 1 January 2027 that becomes open-weight during 2027 qualifies for the end-of-year comparison, even if it was initially proprietary. Models with weights released before the deadline qualify even if the developer subsequently removes the download, provided the release can be independently verified. Distinct officially evaluated configurations of the same underlying model, including reasoning-effort variants, may be treated as separate leaderboard entries where Epoch AI assigns them separate general ECI scores. Each qualifying configuration must be usable with the publicly available model weights and must not depend on undisclosed proprietary model components. Quantized versions, fine-tunes, and other derivatives qualify if Epoch AI reports a separate general ECI score for them and their weights meet the eligibility requirements. For the baseline, models released on 1 January 2027 are included even if their ECI scores are not published until later. Likewise, open-weight models released on 31 December 2027 qualify even if Epoch AI evaluates them during the first three months of 2028. The question does not require an open-weight model to outperform the best proprietary model available at the end of 2027. It only requires an open-weight model to match or exceed the capability level represented by the highest-scoring model available at the beginning of 2027. Epoch AI's ECI methodology and benchmark coverage may evolve. The common-methodology requirement is intended to prevent changes in scoring methodology from producing an artificial YES or NO outcome. Models without an officially published general ECI score by the resolution deadline will not be assigned estimated scores from other benchmarks or third-party leaderboards. The source for the methodology is Epoch AI's official ECI documentation at https://epoch.ai/data/eci-documentation.