Claude Sonnet 5.5 rollout: migrate thinking-off requests and route everyday work by accepted-result cost

Author: Ahab · Date: September 29, 2026

Quick answer

Anthropic released Claude Sonnet 5.5 on September 28, 2026 with the API model ID claude-sonnet-5-5. It is available through Anthropic's platforms and partner clouds, became the default Sonnet model in Claude Code 2.1.284, and entered a gradual GitHub Copilot rollout. Those three entry points do not share one rollout switch: API callers control a model ID and request contract, Claude Code users inherit product defaults, and Copilot organizations also depend on plan eligibility and administrator model policy.

Sonnet 5.5 keeps Sonnet 5 list pricing at $2 per million input tokens, $10 per million output tokens, $2.50 per million cache-write tokens, and $0.20 per million cache-read tokens. Anthropic reports that it is more than 30% faster and can cost up to 30% less per task, but those are vendor results, not IndieSeek reproductions. Migrate with a frozen canary set and compare cost per accepted result—not headline latency or token price alone.

Who this is for

This guide is for teams calling the Claude API, running Claude Code, or enabling Claude models in GitHub Copilot. It separates everyday tasks that can move to Sonnet 5.5 from complex work that may still need Opus 5.5.

What changed—and what did not

Entry point	Confirmed change	Migration boundary
Claude API	New model ID `claude-sonnet-5-5`; 1M-token context; unchanged Sonnet 5 list price	Replay request shapes, especially calls that disabled thinking
Claude Code	2.1.284 adds Sonnet 5.5 and makes it the default Sonnet model on the Anthropic API	Record the resolved model and effort; a product update can change effective behavior
GitHub Copilot	Sonnet 5.5 is generally available to paid plans across supported Copilot surfaces	Rollout is gradual, and Business/Enterprise policy can enable or block it

Anthropic positions Sonnet 5.5 for well-scoped work such as bug fixes and polished documents, while Opus 5.5 remains the stronger option for open-ended judgment. This is vendor guidance, not a universal benchmark result.

The main request-contract change affects applications that ran Sonnet with thinking disabled. Anthropic says those requests must move to the new between_tools setting before switching. Preserve a known-good Sonnet 5 route until that request shape passes, and use the current schema for the provider you call.

A six-step migration workflow

Inventory entry points. List every API route, Claude Code workspace, Copilot editor, coding agent, and cloud provider. Record the current model ID, effort, thinking mode, cache policy, and administrator policy.
Freeze representative tasks. Keep at least one short edit, bug fix, structured response, multi-tool loop, long-context task, refusal case, and ambiguous architecture decision. Preserve acceptance criteria, not just prompts.
Gate thinking-off traffic. Find requests that disable thinking. Update them to the documented between_tools behavior in a branch or canary, then verify text, tool calls, stream parsing, and conversation continuation before changing the model ID.
Normalize effort. Anthropic says Claude Code and its apps default to Medium effort, while Claude Platform defaults to High. Set effort explicitly when possible; otherwise your comparison mixes model quality with a product default.
Canary each platform separately. An API success does not prove Copilot visibility, and a Copilot model-picker entry does not prove enterprise policy enforcement. Capture the resolved model, plan, policy state, provider, latency, tokens, retries, and acceptance result.
Promote with rollback. Shift a bounded traffic percentage, watch accepted-result cost and failure mix, and retain Sonnet 5 or Opus 5.5 as an explicit rollback route.

Measure accepted-result cost

Use the same corpus, effort, tools, cache state, timeout, and reviewer rubric for each candidate:

request_cost = input_tokens × input_rate
             + cache_write_tokens × cache_write_rate
             + cache_read_tokens × cache_read_rate
             + output_tokens × output_rate

accepted_result_cost = (request_cost + retry_cost + tool_cost + review_cost)
                       / accepted_results

Record p50 and p95 latency separately. If Sonnet 5.5 returns faster but creates more retries or review fixes, the accepted-result denominator exposes the regression. If it passes the same rubric with fewer retries, you have local routing evidence—not merely a vendor claim.

Routing decision tree

Is the task well-scoped with an objective acceptance check?
  yes -> canary Sonnet 5.5 at a fixed effort
  no  -> compare Sonnet 5.5 with Opus 5.5 on the same rubric

Does the request disable thinking?
  yes -> migrate to documented between_tools behavior before model cutover

Is the entry point GitHub Copilot?
  yes -> verify plan, gradual availability, admin policy, and resolved model

Did quality, tool integrity, p95 latency, and accepted-result cost pass?
  yes -> promote a bounded percentage and keep rollback
  no  -> keep the current route and diagnose the failed gate

Common mistakes

Calling Sonnet 5.5 a universal Claude-app default when the confirmed defaults and rollout boundaries differ by product.
Comparing Claude Platform at High effort with Claude Code at Medium effort.
Moving a thinking-off request by changing only the model ID.
Treating Copilot GA as proof that every user already sees the model; GitHub says rollout is gradual.
Letting default enterprise enablement stand in for an explicit model-policy review.
Repeating Anthropic's speed or cost-per-task figures as independently reproduced results.
Ignoring Anthropic's documented safety fallback, which may route higher-risk requests to Sonnet 5 while routine development is generally unaffected.

Make your Mac notch useful with SuperNotch—22 native tools for music, clipboard, focus, screenshots, system controls, and more.

FAQ

What is the Claude Sonnet 5.5 API model ID?

Anthropic documents claude-sonnet-5-5. Partner platforms may expose different identifiers or rollout timing, so pin and verify the ID on the provider you use.

Did Sonnet 5.5 get more expensive than Sonnet 5?

No at standard list price. Anthropic lists $2 per million input tokens and $10 per million output tokens, with $2.50 cache writes and $0.20 cache reads. Your actual cost per accepted task still depends on tokens, retries, tools, review, and provider pricing.

Should every coding task move from Opus 5.5 to Sonnet 5.5?

No. Start with well-scoped tasks that have objective checks. Keep an Opus 5.5 comparison route for ambiguous architecture, investigation, and other open-ended work; the separate Opus 5.5 API migration checklist covers its request-contract changes.

Is seeing Sonnet 5.5 in Copilot enough rollout evidence?

No. Verify the target user's plan, model policy, selected surface, and resolved model. Teams using exact model allowlists should also follow the Claude Code model-policy checklist.