Claude Sonnet 5.5 keeps its predecessor’s API price, but promises cheaper finished work

Anthropic released Claude Sonnet 5.5 on September 28 at the same API token rates as Sonnet 5. The company says the new model completes comparable work using fewer tokens and generates output more than 30% faster. Those are Anthropic’s measurements, not a promise that every application’s bill or end-to-end latency will improve by that amount.

The model page lists the Claude API ID as claude-sonnet-5-5, without a date suffix. It is available through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS; the Bedrock ID is anthropic.claude-sonnet-5-5. It is also in Claude products. The model supports a one-million-token context window and 128,000-token maximum output on the standard API.

The price is unchanged; the number of tokens may not be

Anthropic’s pricing documentation confirms $2 per million input tokens and $10 per million output tokens, the same list rates as Sonnet 5. Cache reads are $0.20 per million tokens; five-minute cache writes are $2.50 and one-hour writes $4. Batch API requests get a 50% input/output discount. Those are separate billing paths, not part of the company’s claimed reduction in the cost of a completed task. For a hypothetical job using one million uncached input tokens and 100,000 output tokens, list-price arithmetic is $3 on Sonnet 5.5 versus $6 on Opus 5.5. That comparison excludes cache behavior, batch discounts, tool charges, geography modifiers and differences in token use or task success.

Anthropic says Sonnet 5.5 uses fewer tokens for the same work and costs up to 30% less per task in its testing. It also says output generation is at least 30% faster than Sonnet 5. Neither figure changes posted rates; output speed is not the elapsed time of an agent run waiting on tools. Teams should compare successful task cost and wall-clock time on their own prompts, with a controlled effort setting and tool environment.

That effort setting matters. Anthropic says Claude’s apps and Claude Code default to Medium, whereas the Claude API defaults to High. Lower effort can save tokens and time, but it changes how long the model reasons. The migration guide advises developers to re-run their effort sweep rather than carry over the Sonnet 5 setting unchanged.

Coding gains, with a mixed comparison to Opus

On Anthropic’s published evaluation table, Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, against 10.3% for Sonnet 5 and 66.4% for Opus 5.5. That is a striking result, but the system card puts the Sonnet 5.5 and Opus 5.5 standard errors at about 2.5 and 2.6 percentage points, respectively. The Sonnet run also included safeguard-triggered fallbacks. A four-point lead in this setup is not a clean verdict that Sonnet has replaced Opus for coding.

Anthropic reports 46.2% for Sonnet 5.5 on FrontierCode 1.1 Main, below Opus 5.5’s 54.4%, and 55.5% on CursorBench 4.0, below Opus’s 57.8%. Its GDPval-AA v2.1 score of 1844 nearly matches Opus’s 1846, but Anthropic says Opus remains stronger on open-ended work demanding sustained judgment. These are vendor-published results rather than independent replications; effort levels, harness choices and comparator data vary across evaluations.

A migration is more than changing the model string

For developers moving from Sonnet 5, one breaking change is easy to miss: thinking: {"type": "disabled"} now returns an error. Use between_tools to turn off up-front thinking at Low, Medium or High effort. Forced tool choice using any or a named tool also returns an error; auto is supported. Applications that replay or edit old thinking blocks need to read the migration guidance on conversation binding before switching traffic.

Anthropic is also launching the model with cyber safeguards. Higher-risk requests can fall back to Sonnet 5, so a response in that category may not come from the new model throughout. Routine software development is intended to remain unaffected, but that behavior belongs in a production evaluation alongside cost and latency. For a well-scoped workload, Sonnet 5.5 is a plausible first model to trial at the existing Sonnet rate. Whether it actually saves money depends on the tasks it finishes, not the launch percentage.