Anthropic shipped Claude Sonnet 5.5 on 28 September 2026, the second model in the 5.5 family after Opus 5.5. The company isn’t claiming a leap in general capability, but a 30% faster generation and up to 30% lower cost for everyday work: bug fixes, documents, spreadsheets, well-scoped tasks. Listed pricing stays exactly where it was.
A model built for speed and cost, not a capability leap
Sonnet 5.5 positions itself as the fast, affordable companion to Opus 5.5 within the same family. Anthropic has already flagged Claude Haiku 5.5, aimed at high-volume, cost-sensitive workloads, sketching a three-tier lineup that mirrors the previous generation. The model ships as claude-sonnet-5-5 on the Claude Platform, as well as on AWS, Google Cloud and Microsoft Azure.
| Benchmark | Claude Sonnet 5 | Claude Sonnet 5.5 |
|---|---|---|
| Terminal-Bench 4.0 (agentic) | 10.3% | 70.6% |
| FrontierCode 1.1 (coding, max effort) | 42.4% | 46.2% |
| CursorBench 4.0 | 34.1% | 55.5% |
| OSWorld 2.1 (computer use, partial) | 57.0% | 80.1% |
The benchmarks that matter to developers
The sharpest jump is on Terminal-Bench 4.0, an agentic command-line evaluation, where Sonnet 5.5 climbs from 10.3% to 70.6%. On CursorBench 4.0, which tracks assisted-coding scenarios close to real editor usage, the model reaches 55.5%, just behind Opus 5.5. Several companies cited by Anthropic report concrete production gains: Box reports 2.4x faster execution while using 12% fewer tokens, and Atlassian says its Rovo agents now run up to 30% faster.
curl https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-sonnet-5-5",
"max_tokens": 1024,
"messages": [{"role": "user", "content": "Summarize this changelog in three bullet points."}]
}'
Same pricing, wider availability
Pricing matches Sonnet 5: $2 per million input tokens, $10 per million output tokens, $0.20 for cache reads and $2.50 for cache writes. Anthropic also highlights strengthened safety guardrails at the same level as Opus 5.5, including classifiers meant to curb attempts at distilling the model through third parties.
On CursorBench 4.0, Sonnet 5.5 scores 55.5%, just behind Opus 5.5: a performance-to-cost trade-off built for teams that can’t always afford the top-tier model’s price tag.
What this means for integration
Sonnet 5.5 joined the list of models available in GitHub Copilot on 28 September, followed the next day by a freshly released GPT-6.1 Sol from OpenAI. That near-instant rollout, arriving almost the same day as each vendor’s own announcement, extends questions of billing and governance already raised by previous Copilot policy changes. The model also builds on the same tool-calling foundations via the MCP protocol, whose specification was revised this summer, and fits a broader trend: product teams increasingly looking to hand long-running tasks to agents without constant supervision.
Pricing hasn’t moved, but model behavior has: an automatic switch to the latest available version can change results already validated on a live project, without a single configuration line being touched on the team’s side.
Key takeaways
Claude Sonnet 5.5 isn’t trying to outrank Opus 5.5 on capability; it’s making Sonnet 5’s quality level faster and less token-hungry at the same listed price. The clearest gains show up on agentic and assisted-coding tasks, which explains its near-instant adoption in tools like GitHub Copilot.
One thing worth watching closely: tool vendors are adding these models almost the day they ship, without always leaving time to validate them against one’s own use cases. On client projects, I still prefer pinning the model version in configuration rather than automatically following whatever ships latest: a speed gain is worth nothing if behavior shifts mid-engagement without warning. — Simon Janvier
Primary source: Anthropic, “Introducing Claude Sonnet 5.5”.
