GPT-5.5 Pro vs Claude Opus 5.5

Published configuration: {"variant": "Pro", "third_party_aggregator": true}. ARC-AGI-1 96.5% 92nd percentile · rank 9 of 97 Published model-family result, not a test of TokenRex or its default effort.

Published configuration: {"variant": "Pro", "third_party_aggregator": true}. ARC-AGI-2 84.6% 91st percentile · rank 10 of 99 Published model-family result, not a test of TokenRex or its default effort.

Published configuration: {"variant": "Pro", "third_party_aggregator": true}. BrowseComp 90.1% 88th percentile · rank 8 of 60 Published model-family result, not a test of TokenRex or its default effort.

Published configuration: {"variant": "Pro", "tools": true, "third_party_aggregator": true}. Humanity's Last Exam 57.2% with tools / 43.1% no tools Published model-family result, not a test of TokenRex or its default effort.

Published configuration: {"variant": "Pro", "tools": false, "third_party_aggregator": true}. Humanity's Last Exam 57.2% with tools / 43.1% no tools Published model-family result, not a test of TokenRex or its default effort.

Published configuration: {"variant": "Pro", "third_party_aggregator": true}. FrontierMath 2025-02-28 Private 52.4% 83rd percentile · rank 7 of 37 Published model-family result, not a test of TokenRex or its default effort.

Published configuration: {"variant": "Pro", "third_party_aggregator": true}. FrontierMath Tier 4 2025-07-01 Private 39.6% 78th percentile · rank 6 of 24 Published model-family result, not a test of TokenRex or its default effort.

Published configuration: {"variant": "Pro", "third_party_aggregator": true}. GDPval 82.3% Generalization · 88th percentile · rank 3 of 18 Published model-family result, not a test of TokenRex or its default effort.

Published configuration: {"variant": "Pro", "third_party_aggregator": true}. SimpleBench 76.9% 94th percentile · rank 3 of 36 Published model-family result, not a test of TokenRex or its default effort.

Provider launch evaluation with adaptive thinking at max effort and production safeguards; not the default medium catalog configuration.

Cognition FrontierCode 1.1; Claude Opus 5.5; medium effort; claude-code harness; Main subset (100 tasks). Published weighted rubric score (new_score × 100), not the all-or-nothing pass rate; solution-source violations score zero. Source field new_score=0.5464. Publisher evaluation, not a TokenRex measurement; provider effort configuration is not a guarantee of the API's fixed thinking budget.

Provider launch evaluation with adaptive thinking at max effort and production safeguards; not the default medium catalog configuration.

Row explicitly labeled 'Opus 5.5 Extra High' (xhigh) at 56.0%, $6.98/task. Kept distinct from the existing 57.8% entry, which is Anthropic's own self-reported max-effort figure from the launch page, and from CursorBench's own 'Opus 5.5 Max' row (57.8%, $13.43/task) — same number as Anthropic's but a separately sourced/priced data point.

Provider launch evaluation with adaptive thinking at max effort and production safeguards; not the default medium catalog configuration.

Provider launch evaluation with adaptive thinking at max effort and production safeguards; not the default medium catalog configuration.

Provider launch evaluation with adaptive thinking at max effort and production safeguards; not the default medium catalog configuration.

Explicitly the xhigh effort configuration ('Claude Opus 5.5 (xhigh)'); AA also separately reports 59% at max effort for the same model on this benchmark — kept distinct and not merged. Independent AA harness, not Anthropic's own 58.7% (max effort) figure already on file.

Provider launch evaluation with adaptive thinking at max effort and production safeguards; not the default medium catalog configuration.

Provider launch evaluation with adaptive thinking at max effort and production safeguards; not the default medium catalog configuration.

Launch report explicitly states xhigh effort, standard error ±2.6 points; production safeguards. Not a medium or fast-mode measurement.

Explicitly the Xhigh effort configuration ('Claude Opus 5.5 (Xhigh, Default Fallback)') on AA's independently-run Terminal-Bench 4.0 harness — matches the requested Opus 5.5 xhigh condition exactly, not max/fast/default. Distinct from Anthropic's own self-reported xhigh figure of 66.4% already on file (different methodology/harness) and from AA's narrative max-effort figure of 59.6% quoted in AA's launch article.

Independent AA run, 'Claude Opus 5.5 (Max, Default Fallback)' configuration — AA's max-effort-with-fallback harness, not the catalog medium default. Not comparable to the Xhigh config (56) or High config (54) also reported by AA.

Independent AA evaluation, reported for the top-scoring (max-effort/fallback) AA configuration used for the Intelligence Index; AA states this is the new best score on this eval, ahead of Fable 5.1's prior 59.1%. Effort level for this specific sub-score is implied (Intelligence Index default config), not separately confirmed for every effort tier.

Independent AA evaluation under the Intelligence Index max-effort/fallback configuration; AA reports this surpasses Fable 5.1's prior best of 63.1%.

AA's private frontier knowledge-work evaluation using the open-source Stirrup reference agent harness; max-effort/fallback configuration (the configuration AA headlines for Opus 5.5). Not the Xhigh-specific Briefcase score (1768) also published by AA.

Cognition FrontierCode 1.1; Claude Opus 5.5; medium effort; claude-code harness; Extended subset (150 tasks). Published weighted rubric score (new_score × 100), not the all-or-nothing pass rate; solution-source violations score zero. Source field new_score=0.6527. Publisher evaluation, not a TokenRex measurement; provider effort configuration is not a guarantee of the API's fixed thinking budget.