Most intelligent AI model: Opus 5.5 leads two indexes

24 records on one version of AA’s index

Each step in Figure 1 is the highest v4.3.2 score among models AA dates as released by that day. This is a retrospective record, with all models placed on the same version of the index. It does not reconstruct which model AA ranked first when each release happened.

AA combines ten evaluations: 30% agents, 30% general knowledge, 20% coding and 20% scientific reasoning. Private held-out tests account for 45% of the weight.

OpenAI and Anthropic account for 11 records each; Google and xAI account for one apiece. Claude 3 Opus is the first non-OpenAI record, in March 2024. OpenAI’s last record on this scale is GPT-5.4 in March 2026; every record since April belongs to Anthropic.

Claude Fable 5 sets the largest step, from 41.79 to 49.63 in June 2026: 7.8 points rounded. Its entry includes an Opus 4.8 fallback, which answered about 8% of AA’s tasks.

The first 19 records are estimates

The first measured record is Claude Opus 4.8, released on May 28, 2026, at 41.79. Everything before it carries AA’s estimate flag, and AA publishes no method for those estimates. Even a model measured on an older version can return to estimated status when the index changes.

The flat early years need that qualification. On v4.3.2, Claude 3 Opus keeps the record through GPT-4o and the two Claude 3.5 Sonnets. Epoch places GPT-4o at 128.97 and Claude 3.5 Sonnet at 130.00, above Claude 3 Opus at 126.91; METR’s horizons also increase across these releases.

Small record gaps are not decisive wins. AA quotes a roughly ±1-point 95% interval; four 2025 handovers were by smaller gaps. GPT-5 at 22.98, Claude 4.1 Opus at 22.85 and Grok 4 at 22.49 are especially closely grouped—and all three values are estimates.

Re-scoring changes who appears to have led

Eleven models that led AA at the time fall below the current record line, including Gemini 2.5 Pro, Gemini 3.1 Pro and GPT-5.5. Seven models that now set records were never outright leaders at the time. GPT-5.4, for example, only tied on the displayed rounded score.

GPT-5.5 makes the problem concrete. On v4.0.4, it beat GPT-5.4 by 60.2 to 56.8. On v4.3.2, its measured 38.36 falls below GPT-5.4’s estimated 38.98. The retrospective line therefore skips a model AA had actually ranked first.

Nor can launch-day scores be stitched into one time series. GPT-4 scored 84.4 on AA’s relative index in August 2024, and 6.7 as a v4.3.2 estimate. Fable 5.1 fell from 66 on v4.1.1 to 53 on v4.3 six days later after the rescale. Those score changes reflect a different index.

Two indexes agree; the preference board does not

Epoch fits one capability scale to scores from 60 benchmarks, with an emphasis on maths and knowledge. It also treats OpenAI’s Pro variants as separate models. LMArena records which answer voters prefer, and its default text view has used style-controlled scores since May 16, 2025.

On the October 13 board, Gemini 4 Argon (High) leads the point estimates at 1524.8; Muse Wonder shares the top rank within uncertainty. Opus 5.5 (High) is third at 1508.1. ECI has no entries for Muse Wonder or Gemini 4 Argon, so its agreement with AA has narrower coverage.

Agreement at launch has been unusual. GPT-5 in August 2025 and Gemini 3 Pro in November 2025 were the two launches all four platforms ranked first. In early 2026, METR favoured Opus 4.6 at 12.0 hours while AA and ECI named different leaders. ECI and METR are also partly dependent: METR horizons are one of ECI’s 60 inputs.

The shape of progress depends on the scale

AA’s record rose by 8.2 points in 2024, 15.2 in 2025 and 27.2 in the first nine and a half months of 2026. Epoch’s frontier gains were 16.0, 13.5 and 11.9 over those periods. Its September 2026 analysis found a linear 14 ECI points a year for reasoning models since September 2024. A steeper AA curve alone cannot establish that underlying progress keeps accelerating.

METR uses a different unit: human expert task duration at which an agent succeeds 50% of the time. GPT-4’s 4.0 minutes became 12.0 hours for Opus 4.6, about 180-fold in under three years. The fitted doubling time since 2023 is 128.7 days, with a 104.4–158.0-day interval. The tasks are self-contained software, ML and cybersecurity work, and their error bars are roughly a factor of two.

METR cannot settle the October lead. Its latest covered model is Mythos Preview (early), released in April 2026, with a 17.4-hour horizon above the suite’s reliable 16-hour range. The page has stopped active updates. A straight ECI line and a rising METR curve can still describe similar progress, because their units differ.

Every step, with its evaluation setting

Table 2 lists all 24 records by release date. Reasoning settings and fallback configurations remain part of the model name: Opus 5.5 ranges from 42.31 at Low to 57.62 at Max on AA. A score for one setting is not a score for every configuration of that model.

Availability also matters. Fable 5 was listed unavailable in June 2026; Gemini 4 Argon and Muse Wonder Full were in limited release. These evaluation entries do not guarantee that every reader can access every model.

Dates use AA’s API release convention; AA places o1 on December 5, 2024, while Epoch uses December 17. Gemini 3 Pro’s November 18, 2025 date comes from AA’s launch article because its leaderboard entry has no release date.

How we compiled this

AA’s trends page and leaderboard data provide the v4.3.2 scores and release dates. We keep a record when it exceeds every earlier release’s score, using the highest-scoring entry for each release day. That excludes same-day lower scorers such as o1-mini and Opus 4.6. The table retains the underlying decimal values and AA’s estimate flags.

Cross-checks use Epoch’s ECI CSV, METR’s TH1.1 results, and LMArena’s official history dataset with as-published snapshots. LMArena’s history file omits retired models, so it alone cannot rebuild the earliest leaderboards. AA publishes no human baseline for its index.