Claude Sonnet 5.5 is Anthropic’s current Sonnet model for coding agents, long-running tool workflows and professional AI applications. It keeps the accessible $2 per million input tokens and $10 per million output tokens of Claude Sonnet 5, but it is not a model-ID-only upgrade. Thinking behavior, forced tool use, response blocks and some computer-use integrations changed.
The direct API model ID is claude-sonnet-5-5. Anthropic released the model on September 28, 2026 and lists it as active, with retirement no sooner than September 28, 2027.
Claude Sonnet 5.5 quick specs
| Spec | Claude Sonnet 5.5 |
|---|---|
| API model ID | claude-sonnet-5-5 |
| Release date | September 28, 2026 |
| Status | Active, latest Sonnet generation |
| Context window | 1,000,000 tokens |
| Maximum output | 128,000 tokens |
| Input price | $2 per 1M tokens |
| Output price | $10 per 1M tokens |
| Cache read | $0.20 per 1M tokens |
| 5-minute cache write | $2.50 per 1M tokens |
| 1-hour cache write | $4 per 1M tokens |
| Batch pricing | 50% off standard input and output |
| Input and output | Text and images to text |
| Thinking | Adaptive thinking, high effort by default |
| Knowledge cutoff | June 2026 |
These are Anthropic-direct API specifications. Cloud providers can use different identifiers, regions, release timing and billing conventions.
How much does Claude Sonnet 5.5 cost?
Standard pricing is unchanged from Sonnet 5:
| Billing item | Price per 1M tokens |
|---|---|
| Input | $2.00 |
| Cache read | $0.20 |
| 5-minute cache write | $2.50 |
| 1-hour cache write | $4.00 |
| Output | $10.00 |
Batch processing reduces standard input and output rates by 50%, but it is asynchronous and should not be treated as a latency-equivalent API route. Prompt caching can lower repeated-context cost when system instructions, repositories or evaluation datasets remain stable. Sonnet 5.5 requires at least 512 tokens for a cacheable prompt prefix.
Price per token is only the start of an agent cost model. Tool calls, retries, cache-hit rates and reviewer time affect completed-task cost. Use the AI API pricing comparison and real AI coding cost guide to model those layers.
What changed from Sonnet 5?
Anthropic documents five migration-sensitive areas. Read the dedicated Sonnet 5.5 vs Sonnet 5 comparison before replacing a production model ID.
First, disabling thinking now selects between_tools behavior rather than fully reproducing the older no-thinking flow. This lets the model reason between tool calls at high effort or below. It is not supported with xhigh or max effort, and effort cannot change halfway through a preserved-thinking conversation.
Second, forced tool selection can return an error. Agents that assume tool_choice always forces a valid call need an explicit fallback.
Third, thinking blocks are tied to the model and conversation. Switching models inside a conversation while replaying preserved thinking can fail. Store and replay message history according to Anthropic’s documented rules.
Fourth, Claude API and Google Cloud integrations must use the current computer-use tool instead of the older computer_20251124 version. AWS retains a different compatibility boundary, so provider-specific tests matter.
Fifth, the advisor tool no longer accepts Sonnet 5, Opus 4.7 or Opus 4.8 as advisor models. Do not assume a previously valid advisor configuration will survive the migration.
There is also a response-shape consideration: text between tool calls can appear in thinking blocks. A UI that renders only ordinary text blocks may appear silent while the agent continues working.
Thinking, effort and sampling
Sonnet 5.5 uses adaptive thinking and defaults to high effort. Developers can choose lower or higher effort to trade latency and cost for more deliberate work, but the exact benefit depends on the task.
Non-default temperature, top_p or top_k values are rejected. Remove inherited sampling settings instead of silently assuming they still apply. For production agents, treat the reasoning configuration as part of the tested contract alongside tools, schemas and timeouts.
The AI Testing & Evaluation hub explains how to build regression sets for model changes. Include representative tool failures, long conversations, structured output and recovery paths rather than evaluating only clean prompts.
Tools and coding-agent use cases
Sonnet 5.5 supports tool use, vision, PDF workflows, Files API use, prompt caching and batch processing. It is designed for sustained coding and agent work, but the model still operates inside the permissions supplied by the host product.
Good evaluation targets include:
- repository-scale bug fixing
- terminal and browser workflows
- code review with test execution
- long-document analysis
- multi-step tool orchestration
- structured extraction with schema validation
- professional workflows that need traceable approvals
Use the Coding Agents guide for product-level differences and the AI Security hub for credentials, permissions and sandboxing. A stronger model is not a substitute for least privilege or human approval.
Availability and provider IDs
Anthropic lists Sonnet 5.5 across the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Direct and cloud identifiers differ. The direct model ID is claude-sonnet-5-5; Bedrock uses anthropic.claude-sonnet-5-5.
GitHub also made Sonnet 5.5 generally available in Copilot for eligible paid plans through a gradual rollout. Copilot model access, policy controls and usage billing are separate from Anthropic API availability.
Provider support should be verified at deployment time. A model name appearing in one platform does not prove identical features, regions, cache behavior or retirement policy elsewhere.
Benchmarks, correctly interpreted
Anthropic reports substantial gains on agentic coding and computer-use evaluations. Its launch materials show 70.6% on Terminal-Bench 4 at the stated setting, compared with 10.3% for Sonnet 5 and 66.4% for Opus 5.5 in the same vendor table. Anthropic also reports that Sonnet 5.5 can be more than 30% faster and reduce cost per task by up to 30% versus Sonnet 5.
Those are first-party results, not independent universal rankings. Agent harnesses, effort, tool design and scoring rules can dominate outcomes. Do not combine figures from different tables into one synthetic score. Re-run your own repository and workflow evaluations before changing production traffic.
Sonnet 5.5 versus Opus 5.5 and GPT-6 Sol
Choose Sonnet 5.5 vs Opus 5.5 when the decision is within Anthropic’s family. Sonnet is the lower-priced fast option, while Opus targets the hardest autonomous and knowledge-work tasks at $4/$20.
Choose Sonnet 5.5 vs GPT-6 Sol for a cross-provider decision. Both start at $2 input and $10 output per million tokens, but their long-context pricing, tools, reasoning controls and provider ecosystems differ.
There is no credible universal winner. Compare successful-task cost, latency, tool accuracy, output quality, policy controls and operational fit.
Migration checklist
- Change the model ID in a non-production environment.
- Remove unsupported custom sampling parameters.
- Test thinking and effort behavior at the settings you plan to use.
- Exercise forced-tool, failed-tool and retry paths.
- Verify response-block rendering between tool calls.
- Update computer-use tool versions by provider.
- Test advisor-model configurations.
- Recount token use, cache hits and completed-task cost.
- Run security and permission checks through the AI Application Architecture hub.
- Keep a controlled rollback to Sonnet 5 until production metrics are stable.
My take
Claude Sonnet 5.5 is a real engineering release, not a renamed Sonnet 5. The unchanged token rates make it attractive, while the changed tool and thinking semantics make careless migration risky. New Anthropic integrations should normally start their evaluation with Sonnet 5.5. Existing Sonnet 5 systems should migrate only after their complete agent loop passes regression tests.
Primary sources
- Anthropic Sonnet 5.5 model overview
- Anthropic Sonnet 5.5 announcement
- Anthropic Sonnet 5.5 prompting and migration guidance
Frequently asked questions
What is the Claude Sonnet 5.5 API model ID?
The direct Claude API model ID is claude-sonnet-5-5.
How much does Claude Sonnet 5.5 cost?
Standard pricing is $2 per million input tokens and $10 per million output tokens. Cache reads cost $0.20, 5-minute cache writes cost $2.50 and 1-hour cache writes cost $4 per million tokens.
Does Sonnet 5.5 have a 1M context window?
Yes. Anthropic documents a 1 million token context window and up to 128,000 output tokens.
Is Sonnet 5.5 a drop-in replacement for Sonnet 5?
No. Thinking, forced tool use, response blocks, computer-use tooling and advisor compatibility require testing.
Is Sonnet 5.5 cheaper than Opus 5.5?
Yes by list price. Sonnet costs $2/$10 per million input/output tokens, while Opus costs $4/$20. Completed-task cost still depends on performance and retries.