Claude Sonnet 5.5 Arrives in Copilot: Price, Speed and Migration Guide

Claude Sonnet 5.5 is a new option for everyday coding and document work, including in GitHub Copilot. The useful question is not whether its launch chart looks impressive. It is whether it completes your actual tasks faster, with fewer paid tokens and less repair.

Anthropic launched Claude Sonnet 5.5 on September 28, 2026. It reports output more than 30% faster than Sonnet 5 and up to 30% lower cost per task in its own testing. Those are vendor-reported results, not a discount to the API token rate or a guarantee for every workload. On the same day, GitHub announced Sonnet 5.5 in Copilot, with a gradual rollout across its paid plans.

What Sonnet 5.5 costs

Anthropic lists Sonnet 5.5 at $2 per million input tokens, $10 per million output tokens and $0.10 per million cache-read tokens. These are the same input and output rates as Sonnet 5. Opus 5.5, by comparison, is listed at $4 input and $20 output per million. Cache writes also differ: $2.50 for Sonnet 5.5 and $5 for Opus 5.5. Our Opus 5.5 pricing guide explains why the share of cached context matters in long coding sessions.

“Up to 30% less per task” means Anthropic observed fewer tokens or steps to finish some work. It does not mean every API invoice falls by 30%. For a fair comparison, run the same task set on Sonnet 5 and 5.5, including retries, tool calls and human corrections. Divide total spend by accepted tasks, not by generated answers.

October 7 cache-read price update

Anthropic’s October 7 announcement cuts Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens. Input and output rates remain $2 and $10 per million tokens in the announcement. The 50% reduction applies to cache reads, not the whole invoice.

For example, 10 million cache-read tokens cost $1 at the new rate instead of $2. This calculation excludes other token charges and services. Recalculate savings from your own cached-token share; Anthropic’s estimate of around 20% lower agentic cost is workload-dependent.

Where GitHub Copilot offers it

GitHub says Sonnet 5.5 is generally available for Copilot Pro, Pro+, Max, Business and Enterprise. Its model picker covers Visual Studio Code, Visual Studio, Copilot CLI, the coding agent, the Copilot app, github.com, mobile, JetBrains IDEs, Xcode and Eclipse. GitHub cautions that rollout is gradual, so an eligible account may not see the option immediately.

GitHub says usage is billed at provider list pricing under its usage-based billing arrangement; check the current Copilot model pricing documentation before adopting it widely. Business and Enterprise administrators can manage access through Copilot model policy. A team should verify both the model’s availability and its policy setting, then run a small repository-specific trial. This is a separate change from the Copilot billing and retention policy changes we covered earlier.

How it compares with Opus 5.5

Anthropic positions Sonnet 5.5 for well-scoped work, bug fixes, documents, slides and spreadsheets. It says Opus 5.5 remains stronger on complex, open-ended assignments requiring sustained judgment. That distinction is more useful than claiming one model wins universally. Use Sonnet for a defined implementation or repeatable review; route uncertain architecture and consequential decisions to the stronger model when your own tests justify the extra cost.

Anthropic reports a 70.6% score for Sonnet 5.5 on Terminal-Bench 4.0, versus 10.3% for Sonnet 5 under the company’s reported setup. Benchmark versions, effort settings, harnesses and task selections can change the result. Treat that gap as a reason to test a representative project, not as a prediction that your backlog will improve by the same margin.

Migration checklist for API and Claude Code users

1. Pin the new model identifier. Anthropic lists claude-sonnet-5-5 for its platform; verify the exact identifier and availability in the provider you use.
2. Check thinking settings. If your Sonnet workflow turns thinking off with disabled, Anthropic says to send between_tools instead, at high effort or below.
3. Replace forced tool choice. Requests using tool_choice types any or tool return an error. Anthropic recommends auto with strict tool use.
4. Test streamed tool output. Text between tool calls can arrive inside thinking blocks. Set a display value when using adaptive thinking, or an interface may appear silent between calls.
5. Replay accepted tasks. Compare correctness, elapsed time, uncached and cached tokens, tool calls and required human edits.
6. Keep rollback simple. Route a small portion of traffic first, save failures and retain the previous model until the new route meets your acceptance threshold.

Anthropic’s model documentation lists a 1-million-token context window, a standard maximum output of 128,000 tokens and availability through the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. Availability and controls can still vary by provider and account. If you use agents with external tools, review object-level tool permissions separately; a better model does not grant itself safe authority.

The practical decision

Sonnet 5.5 is a plausible default for bounded coding and knowledge work because its listed token rate stays flat while Anthropic reports better speed and task efficiency. The Copilot rollout makes that comparison easy for many developers. Adopt it where your own accepted-task cost and review time improve; keep Opus or another model for work where its added judgment pays for itself.

Primary sources

• Anthropic: Claude Sonnet 5.5 launch (September 28, 2026).
• Anthropic: Claude Sonnet 5.5 model documentation.
• Anthropic: What’s new and migration checklist.
• GitHub: Claude Sonnet 5.5 in Copilot (September 28, 2026).