Claude Opus 5.5
Anthropic·
Overview
Opus-tier successor to Claude Opus 5, released two months later and ten days after Anthropic's "Pacing the Frontier" proposal. Priced below its predecessor at $4 / $20 per million input/output tokens (cache reads $0.20), with output more than 30% faster than Opus 5. 1M-token context, 128K max output. Thinking can no longer be switched off; default effort is medium. Ships with the same cyber, bio and distillation safeguards as Fable 5.1; Life Sciences Verification Program access at launch, Cyber Verification Program access to follow. Anthropic reports no GPQA or MMLU-Pro figure.
Capabilities and innovations
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| Humanity’s Last Exam (no tools) | 61.4% | third party HLE 61.4% no-tools, measured by Artificial Analysis at max effort (with fallback). Same source as the Opus 5 entry (52.6). Anthropic reports 64.4 no-tools and 67.7 with tools in the Opus 5.5 system card (Table 8.1.A). |
| SWE-bench Pro | 89.9% | vendor SWE-bench Pro 89.9%, Anthropic system card, max effort (Opus 5: 79.2). No independent run yet. |
| Terminal-Bench 2.x | 87.64% | independent Terminal-Bench 2.1 87.64%, measured by Vals AI at high effort, rank 1, counting tasks completed via Anthropic's server-side safeguard fallback as successes (79.77 when counted as failures) (accessed 2026-09-23). |
| Terminal-Bench 4.0 | 65.15% | independent Terminal-Bench 4.0 65.15%, measured by Vals AI at max effort, rank 1 of 42, counting tasks completed via the server-side safeguard fallback (Opus 5, Opus 4.8) as successes (53.54 when the 30 of 198 affected tasks count as failures) (accessed 2026-10-01). Same source as the Sonnet 5.5 entry. Not yet on the official tbench.ai 4.0 board. 4.0 is its own generation and not comparable with the 2.x or 3.0 fields. Replaces 66.4 (vendor). |
| DeepSWE v1.1 | 74.2% | vendor DeepSWE v1.1 74.2, Anthropic system card. Vendor self-report over its own model. |
| MMMU-Pro | 87.7% | third party MMMU-Pro 87.7%, measured by Artificial Analysis at max effort. Same source as the Opus 5 entry (84.7). |
| ARC-AGI-1 | 98.5% | independent ARC-AGI-1 98.5%, verified by ARC Prize (High effort, the best of five variants; Max and XHigh 97.5) (accessed 2026-09-23). |
| ARC-AGI-2 | 93.3% | independent ARC-AGI-2 93.3%, verified by ARC Prize (High effort; XHigh 92.5, Max 91.7). No ARC-AGI-3 result yet (accessed 2026-09-23). |
- GPQA Diamond: independent evaluation pending — Anthropic reports no GPQA value; Vals AI has archived GPQA as saturated and Artificial Analysis lists it as null for this model (checked 2026-09-23).
- MMLU-Pro: independent evaluation pending — Not in the launch table or system card; no Vals AI run yet (checked 2026-09-23).
- Terminal-Bench 3.0: not reported by the vendor — Not on the official tbench.ai Terminal-Bench 3.0 leaderboard (checked 2026-09-23); 4.0 is the current release.
API pricing
$4 per million input tokens, $20 per million output tokens (USD, provider list price, checked 2026-09-23)
Links
More from Anthropic
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.