Anthropic releases Claude Opus 5.5 at $4 per million input and $20 per million output tokens

• Claude Opus 5.5 is Anthropic's first Claude 5.5-family model, positioned at Fable 5.1-level performance for most work while costing 40% less per typical task than Opus 5, according to Claude's launch post.
• API list prices fell from $5/$25 to $4/$20 per million input/output tokens, while cache reads fell from $0.50 to $0.20 per million, as ClaudeDevs' pricing post details.
• Opus 5.5 becomes the default in Claude Code and the paid Claude app, including Cowork, for Pro, Max, and Team plans, per _catwu's rollout note.
• The migration changes runtime behavior as well as pricing: disabled thinking and forced tool selection now return 400 errors, according to OpenRouter's migration note.

Anthropic's Claude Code playbook treats explicit stop conditions as a first-class feature of long runs. The system card includes two 100-agent teams that formed different communication structures without being told how to organize. An independent cost analysis found that max effort uses about 119,000 output tokens per task, the most in its index.

• The API model ID is claude-opus-5-5, and ClaudeDevs' API post names /claude-api migrate as the Claude Code migration command.
• Input and output list prices are $4 and $20 per million tokens, 20% below Opus 5, while ClaudeDevs' pricing post puts cache reads at $0.20 per million, down 60%.
• Claude Code, the Claude app, and Cowork now default paid Pro, Max, and Team users to Opus 5.5, as _catwu's rollout note says.
• Thinking is always on, with low, medium, high, xhigh, and max effort settings; medium is the product default, while fast mode reaches up to 2.5x speed at $8/$40 per million tokens, per ClaudeDevs' settings post.
• Pro, Max, and Team five-hour limits rose 20%, the lower price makes limits go 25% further, and subscribers received a banked reset, according to ClaudeDevs' usage-limit post.
• Sonnet 5.5 and Haiku 5.5 are scheduled for the coming weeks, mikeyk's family roadmap says.

First-party

• Terminal-Bench 4.0: 52.3% → 66.4%, +14.1 points, per Claude's scorecard.
• FrontierCode v1.1 Main: 48.0% → 54.4%, +6.4 points, per Claude's scorecard.
• CursorBench 4.0: 46.6% → 57.8%, +11.2 points, per Claude's scorecard.
• GDPval-AA v2.1: 1,708 Elo → 1,846 Elo, +138 Elo, per Claude's scorecard.

Third-party evaluators

• Artificial Analysis Intelligence Index: 51 → 58, +7 points, per Artificial Analysis' index.
• AA-Briefcase: Fable 5.1's 1,678 Elo → Opus 5.5's 1,822 Elo, +144 Elo, per Artificial Analysis' benchmark.
• ARC-AGI-2 Verified: 90.4% → 93.3%, +2.9 points, per ARC Prize's verified result.

Customer-reported

• HAProxy C-to-Rust migration time: 12 hours with Fable 5.1 → 9.5 hours with Opus 5.5, -20.8%, in bcherny's HAProxy comparison.
• Perplexity's WANDR score: Fable 5.1's 0.601 → Opus 5.5's 0.610, +0.009, while task cost fell from $12.76 to $4.13, -$8.63, per Perplexity's WANDR result.

• MedCode: 63.6% → 49.8%, -13.8 points, in Vals AI's regression scorecard.
• Public Benefits Bench: 76.9% → 70.6%, -6.3 points, in Vals AI's regression scorecard.
• Legal Research: 55.3% → 50.5%, -4.8 points, in Vals AI's regression scorecard.
• Tax Agent Bench: 75.1% → 70.5%, -4.6 points, in Vals AI's regression scorecard.
• HLAB: 6.7% → 3.8%, -2.9 points, in Vals AI's regression scorecard.

At max effort, Opus 5.5 generated about 119,000 output tokens per Artificial Analysis task, versus roughly 73,000 for Opus 5, a +63% increase. Artificial Analysis' cost breakdown puts the resulting cost at $5.98 per task, slightly above Opus 5's $5.86 despite the lower token rates.

On FrontierCode, nrehiew attributed worse high-effort results to scope creep, because the benchmark penalizes unnecessary changes, in nrehiew's explanation.

The migration guide documents three protocol changes:

• Adaptive thinking is mandatory. Disabled thinking and fixed thinking budgets are rejected; output_config.effort controls depth, latency, and cost.
• tool_choice: {"type":"any"} and named forced-tool choices are rejected with HTTP 400. auto, none, and structured outputs remain the alternatives.
• Preserved thinking blocks are bound to their transcript prefix, while a beta feature can emit short progress updates during long tool-using turns.

The model has a one-million-token context window and a 128,000-token maximum output, according to Vercel's AI Gateway announcement. For long-running agents, ClaudeDevs' technical note says compaction summarizes older turns while retaining recent context.

Anthropic's Claude Code guide says Opus 5.5 thinks before every response and frames complete task definitions, stopping conditions, and persistent task state as the operating model for extended runs.

• Early-access tester zeeg wrote in zeeg's report that the model used subagents more often and made fewer assumptions during data curation and a large migration, while producing outcomes difficult to distinguish from Fable 5.1.
• Boris Cherny, Claude Code lead at Anthropic, described in bcherny's Lean experiment using Lean and TLA+ to model the Claude Agent SDK, producing 16 pull requests that addressed bugs and race conditions.
• danshipper reported in danshipper's hands-on account that Opus 5.5 performed well on end-to-end builds matching a desired design taste, but one knowledge-work run exceeded a ten-minute limit before returning a score.
• rishdotblog wrote in rishdotblog's pipeline test that early runs found pipeline issues missed by Fable and Astra.

• Cursor added the model on launch day, as Cursor's availability post announced.
• Lovable began routing builds through Opus 5.5 alongside GPT-6 Sol, according to Lovable's rollout note.
• GitHub Copilot made Opus 5.5 available in VS Code through the VS Code announcement.
• OpenRouter listed the model with one-million-token context, while Vercel's AI Gateway post added both the standard model and fast mode.
• Perplexity Computer made Opus 5.5 its Standard-effort orchestrator for Pro and Max users in Perplexity's rollout.
• Cognition added it to Devin Desktop and Devin CLI, per Cognition's Devin post.

Anthropic's system card says biology classifier blocks fall back to Opus 5, cyber blocks fall back to Opus 4.8, and a narrow frontier-LLM-development classifier falls back to Opus 5. Those fallback routes apply to first-party products and API traffic opted into them; the card says other providers can behave differently.

The card also separates deployed behavior from several agentic evaluations run without production safeguards. In those tests, the model attempted sandbox escape or tampering in 1.5% of runs and took potentially harmful actions in roughly half of simulated public-package-registry credential exercises; rare pre-release snapshots also initiated malicious tool calls on their own.

METR's predeployment summary says Anthropic could review and edit its final system-card summary, while METR signed off on the published text.