Anthropic released Claude Opus 5.5 on September 22, 2026, cutting Claude Platform API pricing from $5/$25 to $4/$20 per million input/output tokens while posting double-digit gains across agentic-coding and knowledge-work benchmarks. Claude Code v2.1.280 and the Anthropic Python SDK v1.8.0 shipped the same day, making Opus 5.5 the new default model and adding MCP tool-list pinning for production agents.
What changed
Claude Opus 5.5 is the first release in Anthropic's 5.5 family, available immediately as claude-opus-5-5 on the Claude Platform API, Claude.ai, Claude Code, and through AWS, Google Cloud, and Microsoft Azure. Standard pricing drops from Opus 5's $5/$25 per million input/output tokens to $4/$20 — a 40% reduction for typical workloads — and cached-read pricing falls 60% to $0.20 per million tokens. A Fast mode, running roughly 2.5x faster, is priced at $8/$40 per million tokens, and standard output generation is itself more than 30% faster than Opus 5.
Benchmark gains follow the same pattern across agentic and knowledge-work tasks. On Terminal-Bench 4.0, Opus 5.5 scores 66.4% against Opus 5's 52.3% and Claude Fable 5.1's 55.8%. It reaches 54.4% on FrontierCode v1.1 and 57.8% on CursorBench 4.0. On the knowledge-work benchmark GDPval-AA v2.1 it posts an Elo of 1846, up from Opus 5's 1708, and the AutomationBench pass rate rises to 40.0% from 26.9%. Additional results include 81.8% partial completion on the computer-use benchmark OSWorld 2.0, 67.7% on Humanity's Last Exam with tool use, and 89.0% on the chart-reading benchmark Chartography with tool use.
Safety work shipped alongside the model: expanded third-party alignment testing by Frontier Design and METR, routing of cybersecurity-related tasks to Opus 4.8, a "preserved thinking" mechanism aimed at resisting distillation, and EU AI Act-compliant watermarking. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow within weeks.
Claude Code v2.1.280 makes Opus 5.5 the default model the same day, switching Pro and Team Standard plans from Sonnet to Opus by default, and fixes several agentic-loop bugs — including auto mode retrying safety-blocked actions indefinitely and a role 'system' must precede an 'assistant' message API error that broke conversations every turn. Anthropic Python SDK v1.8.0 adds beta MCP tool-list pinning, letting applications freeze the tool list an MCP server exposes for the duration of a session.
Who this affects
Teams already budgeting for Opus-class models on agentic coding or long-running tasks get a direct win: same tier, lower price, faster output, and measurably better scores on coding and computer-use benchmarks. Anyone running Claude Code on Pro or Team Standard will notice the default model change immediately, since it now defaults to Opus rather than Sonnet — a cost and latency trade-off worth checking against usage patterns. Developers operating MCP servers in production should look at the new tool-list pinning in the Python SDK, since it closes a class of bugs where a server silently changing its exposed tools could break an in-flight session. Anyone still routing cybersecurity-adjacent tasks through Opus should note Anthropic now routes those specifically to Opus 4.8 rather than 5.5. Teams on Sonnet or Haiku have nothing to act on yet; those tiers arrive in the coming weeks.
Verdict
Upgrade now for agentic-coding and long-running-task workloads: Opus 5.5 costs 40% less than Opus 5 on standard tokens, generates output more than 30% faster, and outscores Opus 5 on every benchmark Anthropic published, including a jump from 52.3% to 66.4% on Terminal-Bench 4.0. There is no scenario where staying on Opus 5 is cheaper or faster, so the switch is close to a pure win for API users already on the Opus tier. Claude Code users on Pro or Team Standard should confirm the automatic default-model switch to Opus fits their budget, since it changes per-session cost even without any action taken. Teams waiting on Sonnet or Haiku pricing parity should hold off migrating other workloads until the 5.5 versions of those tiers ship, which Anthropic says will happen within weeks.
Tracked daily from official release feeds and vendor changelogs. Full archive: https://media.patentllm.org
Top comments (0)