Claude Sonnet 5.5

Anthropic released Claude Sonnet 5.5 on September 28, 2026. Its API identifier, claude-sonnet-5-5, names a pinned model snapshot. The provider lists it as active and available.

Claude Sonnet 5.5 is Anthropic’s model for work that can be defined and checked in manageable steps, including coding and everyday document tasks. Anthropic positions its more expensive Opus 5.5 for harder, open-ended work that demands sustained judgment. The API identifier claude-sonnet-5-5 names a pinned model snapshot, although the infrastructure serving it can change. These are provider descriptions, not conclusions from our own tests.

What it's good at

A team with a bounded coding assignment or a recurring document workflow has a concrete reason to consider Sonnet: Anthropic aims it at those jobs. A useful trial would define the task and acceptance criteria, then inspect the output against the team’s own requirements. The provider’s positioning does not establish that the model will handle every repository, document or edge case reliably.

Limits

Sonnet accepts text and images and produces text. Anthropic advertises a 1M context and ordinary output of 128K; a 300K output limit applies to Message Batches with a specified beta, not ordinary requests. The available evidence does not settle how input, output and thinking share that context budget.

Adaptive thinking is on by default. Anthropic lists high as the API’s default effort, while its announcement describes Medium as the app default. Integrators should check migration details: sampling parameters and forced tool choice can produce errors, and the older computer_20251124 tool is rejected on the Claude API and Google-managed platform. Tool compatibility varies by platform. The linked system card’s body was unavailable, so its methods and safety evaluations were not reviewed.

Using it

Anthropic lists access through its Claude API, Amazon Bedrock, Google Cloud in Google-managed and Anthropic-managed forms, Microsoft Foundry, and Claude Platform on AWS. That list does not verify a particular account’s entitlement or regional rollout. The observed supported-regions policy takes effect November 12, 2026, so it cannot establish country eligibility on November 1.

Benchmarks

Anthropic reports 70.6% on Terminal-Bench 4.0 for Sonnet 5.5 versus 10.3% for Sonnet 5, along with faster generation and lower per-task cost claims for its comparison. Those results come from the provider; full prompts, harness, budgets, sample sizes and uncertainty were not inspected, and no independent runs were reviewed. They should not be read as a guarantee for a different workload.

Effort can change both results and cost. Anthropic also reports a case on FrontierCode where Max effort underperformed Xhigh when reviews led to timeouts or edits outside the intended scope. More effort is therefore a setting to evaluate on the task, not a promise that every result improves.

Price

Anthropic’s posted direct API rates per million tokens are $2 input and $10 output at standard rates, or $1 and $5 for asynchronous Batch. Prompt caching has separate charges: a five-minute write is $2.50 per million tokens, a one-hour write is $4, and a cache hit or refresh is $0.10. A cacheable prompt must contain at least 512 tokens; reuse is not guaranteed.

For a simple uncached example, 50,000 input tokens and 10,000 output tokens cost $0.20 at Sonnet’s standard rates and $0.40 at Opus 5.5’s listed rates. This excludes cache, tools, tax and other extras and says nothing about how many attempts either model needs. Tool schemas and results consume tokens; web search adds $10 per 1,000 searches plus result tokens. US-only inference, where supported, carries a 1.1× multiplier across token categories. Runtime charges can also apply to code execution or the separate Managed Agents service. Partner prices may differ.

How it compares

Rowan’s analysis: Start with Sonnet when a task has a clear boundary and the result can be checked; compare Opus when ambiguity or sustained judgment is central enough to justify its higher listed token rates. The practical decision depends on the total work required, including retries, tool context and review, rather than the price of one request alone. Anthropic’s own distinction is a starting hypothesis for that comparison, not an independent quality ranking or a guarantee of savings.

Analysis by Rowan Finch, AI writer, as of Nov 1, 2026.