Claude Opus 5.5
Released Sep 22, 2026 · knowledge cutoff 2026-06
- Status
- Frontier
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1M
- Max output
- 128K
- Speed (medium effort)
- 75 tok/s · 22s to first answer token
- Price ($/MTok in / out)
- $4 / $20
- Cost per task (max effort)
- $5.98
- Open weights
- No
Benchmarks4 of 10 reported
SWE-bench Pro
Terminal-Bench 2.1
Humanity's Last Exam
ARC-AGI-2
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
AIME (latest)retired
LMArena Elo
GDPval-AA v2
Strengths & weaknesses
Strengths
- Finishes the same coding task in fewer steps and tokens than Opus 5 — its default medium effort matches Opus 5 at high, so most of the gain comes without paying for max
- Reads dense charts, flowcharts and screenshots accurately without crop-and-zoom tooling, so vision scaffolding built for Opus 5 can usually be dropped
- Answers are shorter and lead with the conclusion — the bold-and-bullet 'Claudish' scaffolding Opus 5 leaned on is largely gone
- Holds up over multi-hour unattended runs and resists indirect prompt injection better than any earlier Opus
Weaknesses
- Thinks longer per turn than Opus 5 at the same effort label, and at max it can burn the whole 128K output budget before answering — reserve xhigh and max for measured gains
- Stops to report progress partway through long tasks, ending the turn with text rather than a tool call — unattended loops need a checklist or re-prompt harness to keep it going
- Code review comes back with more findings and roughly half again the tokens of Opus 5 — the extra flags still need a human to triage
HLE 64.4% no tools / 67.7% with tools, SWE-bench Pro 89.9% and GDPval-AA v2.1 1846 are Anthropic's system-card figures (Table 8.1.A: adaptive thinking at max effort, five-trial average, production safeguards on — cyber tasks routed to Opus 4.8 and bio/frontier-LLM tasks to Opus 5 where they intervened); GDPval-AA stays null here because the tracked key is the v2 scale. Anthropic's launch table leads with Terminal-Bench 4.0 66.4% (xhigh; not comparable with 2.1), FrontierCode v1.1 54.4%, CursorBench 4.0 57.8%, OSWorld 2.0 81.8% partial / 48.7% strict, Terminal-Bench-Science 0.1 58.7% and AutomationBench 40.0%. Terminal-Bench 2.1 87.6% is Vals AI's independent-harness measurement at high effort (vals.ai/models/anthropic_claude-opus-5-5) — third-party, and 26 of 267 tasks needed server-side fallbacks, so it drops to 79.8% if those count as failures. ARC-AGI-2 91.7% is ARC Prize-verified at max effort (arcprize.org/results/anthropic-claude-opus-5-5, 2026-09-22); high effort scores 93.3%, xhigh 92.5%, medium 87.5%, low 70.1%. Anthropic no longer reports SWE-bench Verified, GPQA Diamond, MMLU-Pro or AIME, and the model is not on arena.ai's text leaderboard as of its 2026-09-13 update. Cost per task is Artificial Analysis' max-effort variant, the only one it lists (Intelligence Index 58, first place); AA had no speed measurement at launch. Default effort is medium (Opus 5 defaulted to high), adaptive thinking cannot be disabled, and forced tool_choice is rejected. Cache reads are $0.20/MTok (0.05x input); batch $2/$10; fast mode $8/$40 on the first-party API only; max output rises to 300K via the Message Batches API beta. Also served on Microsoft Foundry as claude-opus-5-5.
For developers
API model strings
- anthropic
claude-opus-5-5 - aws-bedrock
anthropic.claude-opus-5-5 - google-vertex
claude-opus-5-5
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor