Two years pass with no new record on Epoch's Capabilities Index

Where things stand (14 Oct 2026)

I downloaded Epoch's general ECI file (https://epoch.ai/data/eci_scores.csv, 274 models, updated 13 Oct 2026). I grouped models by Epoch release date and rebuilt the record frontier from that one data vintage.

• There are 23 record-setting dates from Feb 2023 to Sep 2026. The 22 gaps between them have a median of 41 days and a mean of 59 days.
• The longest gap ever is 317 days: GPT-4 (Mar 2023) to GPT-4 Turbo (Jan 2024). Since mid-2024, no gap has exceeded 104 days.
• 2026 records so far: 5 Feb, 5 Mar, 23 Apr, 24 Jul, 1 Sep, 3 Sep and 22 Sep.
• The current record is Claude Opus 5.5 at 167.33 (22 Sep 2026). GPT-6.1 Sol (166.09, 29 Sep) did not beat it.
• Gemini 4 Argon launched publicly on 13 Oct and is not yet scored.

What this means for timing. The quiet stretch has to start on or after the last record. So the earliest possible end date is 22 Sep 2028, and in practice later. I set every quarter-end through mid-2028 at 0.

Reference class

Neither the ECI history nor the wider modern deep-learning era has ever had a 24-month stall in frontier capability. If records keep arriving at the observed pace (about one per 59 days, treated as random arrivals), the chance of a 730-day gap is about e^-12, which is effectively zero. So any probability comes almost entirely from a regime change.

Paths to YES

The rule is strict: any model from any developer, by any margin, breaks the streak. That leaves a few routes.

1. A binding global freeze on frontier releases lasting at least 2 years. The environment has shifted a lot:

• rogue-agent incidents, including the OpenAI–Hugging Face incident;
• Dario Amodei's essay "We Must Pace the Frontier" (12 Sep 2026), with Altman, Musk and Hassabis agreeing;
• a US–China "Super Intelligence Dialogue";
• polls showing most Americans think AI is moving too fast;
• Manifold prices: international training pause in effect on 1 Jan 2028 at 36%; a 6+ month pause by 2030 at 39%; Ban Superintelligence Act passing by 2027 at 18%.

But several things cut against it:

• Amodei says pacing "does not mean halting," and calls a full pause "unlikely… any time soon."
• Trump calls the doom talk a "HOAX," and the White House Accord is voluntary and has no teeth (CFR).
• Beijing "will not pace the frontier" (SemiAnalysis); Meta rejects slowing down; a federal antitrust suit challenges coordination.
• Labs kept shipping after the essay.
• Most importantly, open and Chinese models trail the closed frontier by only about 4 months, about 8 ECI points (Epoch, May 2026). A freeze limited to US labs would be overtaken quickly, so China and every other developer would have to stop too.

A training pause also lets labs keep releasing improved or already-trained models. Small noisy gains in point estimates would still count as records. I put this path at roughly 0.6–1%.

2. Other routes, about 1.5% combined:

• a catastrophe or war that halts releases while Epoch keeps publishing;
• frontier models moving to restricted access while Epoch leaves them out of the general index. Mythos Preview and Mythos 5 appear only in the Cyber ECI. Open models would still need to stall too;
• an artifact from saturated benchmarks producing ties. GPT-6.1 Sol already scores 100% on FrontierMath Tier 4;
• a genuine plateau, or unknown unknowns.

Each route also needs Epoch to keep the general ECI, or an acknowledged continuation of it, running and publish data covering the end date. If the index stops updating, that cannot resolve YES.

How I built the series

I estimated the probability that the last record before a 24-month gap falls in each quarter from Q4 2026 to Q4 2029:

• about 0.08% for Q4 2026;
• 0.12–0.22% per quarter through 2028;
• 0.28–0.32% per quarter in 2029, reflecting a possible pause-minded administration after the 2028 election.

That totals about 2.7%. Each start is mapped to resolution 8 quarters later (70%) or 9 quarters later (30%, to allow for publication lag). Accumulating gives about 2.6% by end-2031.

Strongest case against my estimate

Taken at face value, the Manifold pause markets would imply roughly 5–10%. I discount them heavily for three reasons:

• they are thin and long-dated;
• a pause on training is not the same as no ECI record from any developer for 24 months;
• China's stance and the 4-month lag of open models make a global zero-progress window much harder than a "pause."

Going the other way, the strict any-increment rule could justify a figure under 1%.

Sources: Epoch ECI data and methodology (epoch.ai/eci; epoch.ai/data/eci-documentation/methodology); Epoch "Open models lag… by 4 months" (29 May 2026); darioamodei.com/post/we-must-pace-the-frontier (12 Sep 2026); TechCrunch (12 Sep 2026); Fortune (15 Sep 2026); CFR on the White House Accord (Sep 2026); SemiAnalysis "Beijing Will Not Pace the Frontier" (Oct 2026); American Action Forum on the antitrust suit (7 Oct 2026); shattered.io on the launch streak (2 Oct 2026); Reuters on the Gemini 4 Argon launch (13 Oct 2026); Manifold market API (pulled 14 Oct 2026).

Question. By the last day of each quarter, have 24 consecutive months passed without any model setting a new highest score on Epoch AI's Capabilities Index?

Definitions.

• Index: Epoch AI’s general Epoch Capabilities Index (ECI), published at epoch.ai/eci, including its downloadable model-score and release-date data. Domain-specific ECIs, individual benchmarks, and filtered leaderboards do not count. Models from all developers are included.
• New highest score: a model’s general ECI point estimate is strictly greater than the point estimates of every model with an earlier Epoch-listed release date. Models released on the same date are treated as a group: that date sets a record if the group’s maximum exceeds the previous maximum. Ties do not set records. Use the most precise official downloadable estimates available, rather than rounded chart labels; overlapping confidence intervals do not prevent a record.
• History and revisions: reconstruct the release-date frontier using one complete official data vintage available at the time of assessment, comparing all models within that vintage. Do not compare numerical scores across vintages. Incorporate Epoch’s historical corrections, backfilled models, and refitted estimates before resolution; a numerical increase caused solely by rescaling or refitting is not a new record dated to the update. Use Epoch’s release dates, not the dates models were first added to the index.
• 24 consecutive months: an interval from a date S through its second calendar anniversary T, with no record-setting release date in (S, T]. There must be an established frontier at S, and T must be after the question’s opening date. S may precede the opening date. If the anniversary date does not exist, use the last day of that month.

Resolves YES if a dated Epoch AI publication or data update, published within the Resolution window and on or before the quarter-end, establishes a qualifying 24-month interval. The publication must cover the index history through at least T. A later publication may establish an earlier qualifying interval, but resolution occurs only on publication, not retrospectively at T. The Observation period convention applies.

An official successor or renamed version counts only if Epoch identifies it as a continuation of the general ECI and supplies a comparable historical series covering the interval and the preceding frontier. Apply its full historical vintage rather than splicing incompatible series. Once YES, later revisions do not reverse resolution.

Resolves NO / Exclusions. Resolve NO for a quarter if no qualifying publication has established the interval by its end. A plateau in a domain-specific or filtered index does not count. An unchanged, abandoned, or inaccessible webpage is not evidence that the interval has elapsed: silence or cessation of index publication alone cannot establish YES. A terminal Epoch publication can establish YES only if its history covers the full qualifying interval. An interval containing any strictly higher record fails, however small the increase or brief the model’s lead.

Resolving source. Epoch AI only: its general ECI leaderboard and downloadable data at epoch.ai/eci, or an Epoch-authored historical analysis explicitly establishing the interval under these definitions. Archived copies may establish what Epoch published and when. Methodology is documented at ECI Documentation and its Methodology page. These are narrower resolving sources than Public confirmation (A2). Apply Resolution date (A3), the Resolution window, and the Observation period conventions.