Anthropic Ships Claude Sonnet 5.5: 30% Faster, Up to 30% Cheaper Per Task
AI寒武纪 · wechat · 2026-09-29
Anthropic released Claude Sonnet 5.5, the fast, low-cost companion to Opus 5.5 for well-scoped everyday tasks, bug fixing, and document/slide/spreadsheet work.
- Over 30% faster than Sonnet 5 with up to 30% lower cost per task at the same price ($2/$10 per M input/output tokens, $0.20 cached reads).
- Scores 70.6% on Terminal-Bench 4.0 (vs 10.3% for Sonnet 5) and is the first Sonnet to beat Pokémon Red from screenshots alone; at Max effort it matches Opus 5.5 on several benchmarks.
- Coding: +10 points over Sonnet 5 on FrontierCode at High effort at 1/15th the cost per task; merges tool calls more aggressively.
- Knowledge work: within 2 points of Opus 5.5 on GDPval-AA; an expert-reviewed 10-page earnings deck draft was deemed ready to send.
- First Sonnet shipping with an anti-distillation safety classifier and expanded retained thinking.
- Live on AWS, Google Cloud, and Azure as claude-sonnet-5-5, with zero data retention; migrating from disabled-thinking requires the new betweentools setting.
Related event: Anthropic Launches Claude Sonnet 5.5: Over 30% Faster, Up to 30% Cheaper(58 posts)→
More from coding & agent
- AI prototype unreliable? Scope it down: let the agent do less, then grow — brandon_galang · 2026-11-13
- Porting an Unreal game to iPhone and iPad with Claude Code, no manual Xcode work — chongdashu · 2026-11-13
- Influzer launches an instant MCP server inspector: paste any URL to see its tools — Fine_Airline_6832 · 2026-11-13
- Open-source getmcp-cli turns 1,280-endpoint OpenAPI specs into 25-tool MCP servers, cutting tool tokens from 263k to 6k — Worried-Barracuda845 · 2026-11-13
- Who's paying for unused AI coding seats? Per-seat billing is quietly leaking money — rcleffke · 2026-11-13
- Formal verification is the Rust of software correctness — LLMs are making people write specs — arpit_bhayani · 2026-11-13