Anthropic releases Claude Sonnet 5.5 at unchanged token prices

• Claude Sonnet 5.5 shipped across Claude and the API; Anthropic's announcement reports over 30% faster output generation and up to 30% lower cost per task than Sonnet 5.
• Token prices remain $2 per million input and $10 per million output, with $0.20 cache reads, according to Claude's developer update.
• Agentic coding made the biggest jump in Anthropic's table: Terminal-Bench 4.0 rose from 10.3% to 70.6%, +60.3 points, as the launch-table screenshot shows.
• Max effort can erase the savings: Artificial Analysis' independent evaluation measured $7.60 per task, roughly 50% more than Sonnet 5 on its Intelligence Index.
• The API migration changes tool selection and thinking behavior; Claude's building guide documents five breaking changes beyond the model ID swap.

Anthropic's FrontierCode footnote explains a rare benchmark dip at Max effort: extra review subagents sometimes timed out or changed too much. The migration guide says forced tool selection now returns a 400, while Artificial Analysis measured 193,000 output tokens per task at Max. Simon Willison's SVG test consumed the full 128,000-token output allowance without producing the requested pelican.

• Model and access: Sonnet 5.5 is the second Claude 5.5 model after Opus; Haiku 5.5 is due in the coming weeks, per Anthropic's rollout post. Sonnet 5.5 powers Claude's free chat tier, according to Simon Willison's free-tier report; his clarification says Claude Code still requires a paid account.
• API and clouds: The model slug is claude-sonnet-5-5 on the Claude API, Claude Platform on AWS, Google Cloud, and Microsoft Foundry; Bedrock uses anthropic.claude-sonnet-5-5, as the developer guide specifies.
• Price: Input/output rates stay at $2/$10 per million tokens; 5-minute cache writes cost $2.50 per million, and cache reads cost $0.20, per Claude's pricing post. Anthropic attributes its up-to-30% task savings to fewer tokens, while its speed-and-cost post puts output generation more than 30% faster.
• Defaults: Claude apps and Claude Code start at Medium effort, while the Platform defaults to High, according to Claude's developer note. the Claude Code 2.1.284 changelog says Sonnet 5.5 replaced Sonnet 5 as the default Sonnet selection, not the overall Opus default.
• Usage reset: Pro, Max, and Team accounts received a reset that can be applied before October 22, according to Claude's rollout note.

The launch table mixes model effort settings; its Opus Terminal-Bench comparator is Opus's best Xhigh run. Scores below compare Sonnet 5 with Sonnet 5.5 within each source's own evaluation.

First-party

• Terminal-Bench 4.0: 10.3% → 70.6%, +60.3 points, per the launch-table screenshot.
• CursorBench 4.0: 34.1% → 55.5%, +21.4 points, per the benchmark table.
• FrontierCode 1.1 Main at the reported best settings: 42.4% → 52.1% at Xhigh, +9.7 points, per Anthropic's table.
• GDPval-AA v2.1: 1,449 → 1,844 Elo, +395 Elo, per the launch numbers.
• OSWorld 2.1 partial: 57.0% → 80.1%, +23.1 points, per the comparison table.
• Chartography without tools: 15.6% → 61.6%, +46.0 points, per the published chart.

Third-party evaluators

• Artificial Analysis Intelligence Index at Max: 38 → 56, +18 points, per Artificial Analysis' release analysis.
• Artificial Analysis' Terminal-Bench 4.0: 14.1% → 63.6%, +49.5 points, per its independent runs. Its absolute score differs from Anthropic's 70.6% result; the two organizations report separate runs.
• Vals Index accuracy: 59.61% → 69.22%, +9.61 points, per Vals AI's evaluation thread.

Customer-reported

• Box's complex-work eval: 61% → 65%, +4 points, per Box's early-access results; its financial-services subset rose 63% → 81%, +18 points, in the same tests.
• Cognition's FrontierCode 1.1 run in Devin: 56.2% → 64.4%, +8.2 points, per Cognition's Devin launch thread. This is Cognition's agent evaluation, rather than the score in Anthropic's launch table.

More effort sometimes produced more work and a worse result. Anthropic's FrontierCode footnote reports Sonnet 5.5 at Xhigh scoring 52.1% versus 46.2% at Max, −5.9 points; in two cases Cognition examined, extra code-review subagents caused a timeout or out-of-scope edits.

• Vals measured CyberBench at 62% → 60%, −2 points, and Harvey Legal Agent at 5% → 3%, −2 points, in its 18-benchmark thread. The other 16 benchmarks rose.
• Artificial Analysis found factual accuracy of 54% for Sonnet 5.5 versus 66% for Opus 5.5 on AA-Omniscience, −12 points relative to Opus, according to its evaluation. It also reports lower hallucination rates for Sonnet on that test, 47% versus 59%.
• Anthropic says tasks can cost up to 30% less than Sonnet 5 in its own testing, while Artificial Analysis' Max-effort runs cost $7.60 per Intelligence Index task, about 50% more than Sonnet 5 and above Opus 5.5's $5.98. The measurements cover different workloads and effort settings; flat token rates do not fix the task bill.

At Max, Artificial Analysis counted about 193,000 output tokens per task, versus 119,000 for Opus 5.5. Vals also reported occasional 1-million-token context exhaustion and sandbox crashes from too many processes. Simon Willison's hands-on test hit the 128,000-token output limit without an SVG at Max; Xhigh produced one in 41 seconds for 5.74 cents.

Artificial Analysis tested a pre-release deployment with a structured-outputs bug. Anthropic says it fixed the bug for public release, and Artificial Analysis said it would rerun affected evaluations.

The developer model sheet lists a native 1-million-token context window, 128,000 maximum output tokens, five effort levels (low, medium, high, xhigh, max), and the same tokenizer as Sonnet 5. Message Batches can reach 300,000 output tokens with a beta header; the minimum cacheable prompt drops from 1,024 to 512 tokens. A 2000×1500 image costs roughly 2.5 times as many input tokens as on Sonnet 4.6, Sonnet 4.5, or Haiku 4.5 because Sonnet 5.5 uses the high-resolution image tier.

The Sonnet 5.5 migration guide spells out the Messages API changes behind Claude's short developer post:

1. Thinking: Adaptive thinking remains on when the field is omitted. Sonnet 5's thinking.type: "disabled" now returns a 400; between_tools removes up-front thinking at Low, Medium, or High, but returns a 400 at Xhigh or Max. Thinking counts toward billed output and max_tokens.
2. Tool choice: Forced tool_choice: "tool" or "any" returns a 400. auto plus strict: true constrains tool inputs without requiring a tool call; Bedrock lacks strict-tool support for this model.
3. Conversation state: Responses can start with a thinking block, so clients reading content[0].text break. Thinking blocks must travel unchanged through tool loops; progress between tool calls can arrive in those blocks rather than text.
4. Computer use and advisors: The Claude API and Google Cloud require computer_toolset_20260801 instead of computer_20251124; Bedrock still accepts the older tool. Sonnet 5.5 also rejects Opus 4.8, Opus 4.7, and Sonnet 5 as advisors, and returns accepted advisors' advice encrypted.
5. Preserved thinking: Sonnet 5.5 can read Sonnet 5's thinking blocks, but older models cannot read Sonnet 5.5's. Anthropic says thinking remains bound to its originating organization; switching accounts mid-session makes Claude reread the session and generate fresh thinking, per the developer announcement.

Sonnet 5.5 is the first Sonnet shipped with cyber safeguards and fallback routing for higher-risk requests, according to Anthropic's safety note. The release details say flagged cyber requests visibly fall back to Sonnet 5, while routine software development is unaffected; Artificial Analysis observed fallback on roughly 0.1% of its Intelligence Index tasks.

After a day of use, kunchenguid found Opus better at planning ambiguous product features and Sonnet faster at implementation and orchestration in their first-day report. Their working split was Opus for plans, Sonnet for implementation, and the forthcoming Haiku for fixes.

• A small internal-library update went from prompt to PR in about 20 seconds, according to rishdotblog's first run.
• Sonnet handled revisions better than Astra in Dan Shipper's work tests, though Astra remained his favorite overall writing model; quick coding and design iteration favored Sonnet for his colleague, per Shipper's hands-on assessment.
• A photo-to-code painting test had the model write a Python brush engine, render its work, inspect it, and revise it; Lance Martin posted the Sonnet 5 comparison.
• An interactive three.js plasma globe felt close to Opus at half the token price to one builder who made it. Another side-by-side voxel test took Sonnet 5.5 at High 49 minutes and $9.23, versus 22 minutes and $4.85 for Sonnet 5 at High, despite the richer output.

Cursor reported that Sonnet 5.5 performs on par with Opus on many of its tasks in its launch post. Other day-one surfaces span coding agents, gateways, and browser tools:

• Coding agents: Devin Desktop and CLI added Sonnet 5.5, per Cognition's rollout thread; Factory's first tests say High effort checked the actual requirement rather than merely the nearest failing test. Warp's announcement covers both Warp and its Agent CLI.
• Gateways: OpenRouter's launch lists a 1-million-token context window at $2/$10 per million input/output tokens, while Vercel's AI Gateway release exposes anthropic/claude-sonnet-5.5.
• App builders: v0's release added the model, and Julius's rollout reports stronger coding and clearer writing inside its analysis, presentation, and website workflows.
• Browser agents: Hyperbrowser's launch put Sonnet 5.5 in its Agents Playground for navigation, clicks, and typing.