Claim · agents

GPT-5 Codexon Terminal-Bench 2

41.3 to 44.3 the range across every scoreable measurement we hold

3.0 points between the highest and the lowest

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

3 declared configurations, from 1 runner

0 / 3 rows whose cited page still shows this figure, refetched and checked by us

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.

3 of 3 cited pages no longer show the figure we publish. The page was refetched and split at its own structural boundaries, and no block that names this model carries this number. That does not make the measurement wrong: a leaderboard is a living page and a figure recorded when it was there can be gone by the time you follow the link. It does mean the citation, as printed, no longer supports the number, and this is the only place we know of that says so about its own sources.

What the vendor said, and what someone else measured

Matched on which variant of the benchmark was sat, and on nothing else: the vendor ran it on their harness and the independent party on theirs, and that difference is the comparison rather than an obstacle to it. A vendor number with no independent run of the same variant carries no delta, because a number that cannot be compared honestly must not be compared.

Variant	Vendor claims	Independently measured	Difference
as published, no variant stated	44.3	43.4 from 3 runs	+0.9

Every measurement

6 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
44.3 / 100 as published: 44.3%	UNATTRIBUTED shown, never scored	harness Codex CLI	Terminal-Bench via Epoch AI 2025-11-04 CC-BY-4.0	NO_URL2026-08-30 benchmark leaderboard the leaderboard for this benchmark, not a link to this row
44.3 / 100 as published: 44.3%	INDEPENDENT counts toward the ranking	harness Codex CLI	Terminal-Bench v2 Leaderboard via Epoch AI 2025-11-04 CC-BY-4.0	checked: this figure is not on the page today2026-08-30 Terminal-Bench v2 Leaderboard
43.4 / 100 as published: 43.4%	UNATTRIBUTED shown, never scored	harness Terminus 2	Terminal-Bench via Epoch AI 2025-10-31 CC-BY-4.0	NO_URL2026-08-30 benchmark leaderboard the leaderboard for this benchmark, not a link to this row
43.4 / 100 as published: 43.4%	INDEPENDENT counts toward the ranking	harness Terminus 2	Terminal-Bench v2 Leaderboard via Epoch AI 2025-10-31 CC-BY-4.0	checked: this figure is not on the page today2026-08-30 Terminal-Bench v2 Leaderboard
41.3 / 100 as published: 41.3%	UNATTRIBUTED shown, never scored	harness Mini-SWE-Agent	Terminal-Bench via Epoch AI 2025-11-03 CC-BY-4.0	NO_URL2026-08-30 benchmark leaderboard the leaderboard for this benchmark, not a link to this row
41.3 / 100 as published: 41.3%	INDEPENDENT counts toward the ranking	harness Mini-SWE-Agent	Terminal-Bench v2 Leaderboard via Epoch AI 2025-11-03 CC-BY-4.0	checked: this figure is not on the page today2026-08-30 Terminal-Bench v2 Leaderboard

← 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.