Claim · reasoning

Claude Opus 5on ARC-AGI-2

88.3 to 90.4 the range across every scoreable measurement we hold

2.1 points between the highest and the lowest

88.3 the number the board uses: the lower median, a real row with its own provenance

2 declared configurations, from 1 runner

All 3 scoreable measurements here come from ONE runner, so this is not a reconciliation and is not presented as one. What it shows is how far a single runner's number moves when the model is run differently.

Two measurements, one declared configuration, +2.1 points apart. 2 rows here agree on everything the record states, the system identity included, and disagree on the number: 90.4, 88.3. Something moved the score and no source we hold says what. The board resolves it by taking the lower median, which is a rule and not an explanation, so it is named here rather than quietly settled. This happens in 14 of 321 claim pages.

Every measurement

3 rows. Best first. the board takes the LOWER MEDIAN of the scoreable measurements, which is a real row with its own interval and provenance rather than an interpolated average. a vendor self-report is shown beside the independent numbers and is never one of them; it cannot move a rank.

Score	Provenance	Configuration	Who ran it	Where to check it
90.4 / 100 as published: 90.4%	INDEPENDENT counts toward the ranking	effort max	ARC Prize via Epoch AI CC-BY-4.0	NO_URL2026-08-30 benchmark leaderboard the leaderboard for this benchmark, not a link to this row
88.3 / 100 as published: 88.3%	INDEPENDENT counts toward the ranking	effort max	ARC Prize via Epoch AI CC-BY-4.0	NO_URL2026-08-30 benchmark leaderboard the leaderboard for this benchmark, not a link to this row
88.3 / 100 as published: 88.3%	INDEPENDENT counts toward the ranking	effort high	ARC Prize via Epoch AI CC-BY-4.0	NO_URL2026-08-30 benchmark leaderboard the leaderboard for this benchmark, not a link to this row

← The Forge Rankings

Snapshot 2026-08-30-62e4e85f043b, manifest dc46131315046f1e, as of 2026-08-30. where only one party has measured a model on a benchmark, this page says so rather than presenting a single number as a reconciliation.