What is Claude Sonnet 5.5?
Claude Sonnet 5.5 is Anthropic’s active general-purpose language model for applications that need a combination of coding ability, reasoning, speed, and operational flexibility. Anthropic positions it between the higher-capability Claude Opus 5.5 and the lower-cost Haiku line. That makes Sonnet 5.5 a practical middle option: it is intended to handle demanding everyday work while remaining faster and less expensive than the provider’s top-tier model.
The model is aimed at software development, codebase maintenance, long-running coding tasks, professional documents, spreadsheets, presentations, image understanding, and tool-based workflows. It can analyze large amounts of material in one request and can work with external tools when an application supplies them.
According to Anthropic’s supplied model documentation, Claude Sonnet 5.5 was released on September 28, 2026. Its primary Claude API identifier is claude-sonnet-5-5. It is also available through Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. On Amazon Bedrock, the model identifier is anthropic.claude-sonnet-5-5.
Specifications at a glance
| Specification | Claude Sonnet 5.5 |
|---|---|
| Provider | Anthropic |
| Status | Active and listed as latest |
| Context window | 1,000,000 tokens |
| Standard maximum output | 128,000 tokens |
| Batch maximum output | Up to 300,000 tokens with the applicable beta configuration |
| Knowledge cutoff | June 2026 |
| Input | Text and images |
| Output | Text |
| Reasoning | Adaptive thinking with configurable effort |
| Latency positioning | Fast, according to Anthropic’s model information |
The context window is the amount of information the model can consider in a request and its surrounding conversation. A 1 million-token limit is useful for large codebases, lengthy document collections, transcripts, and other tasks where splitting the source material into many smaller requests would be inconvenient. The context limit should not be confused with the maximum response size: a normal request can return up to 128,000 tokens, while certain Batch API configurations can reach 300,000 output tokens.
Pricing and usage costs
On the Claude API, Anthropic lists the primary token prices as:
- Input: $2 per million tokens
- Output: $10 per million tokens
Input tokens are the text, images, and other request content sent to the model. Output tokens are the generated response. The difference matters for applications that repeatedly send large prompts or produce long responses: a workflow that generates extensive code or documents can incur substantially more output cost than a short-answer application.
Claude Sonnet 5.5 also supports prompt caching. A five-minute cache write costs $2.50 per million tokens, while a one-hour cache write costs $4 per million tokens. Cache reads cost $0.20 per million tokens. The minimum cacheable prompt length is 512 tokens. Caching can reduce the cost of repeatedly sending the same large instructions, reference documents, or codebase context, although developers need to account for the separate write and read pricing.
The Batch API applies a 50% discount to input and output pricing. Batch processing is intended for requests that do not require immediate responses, such as bulk classification, document processing, offline code analysis, or large-scale content transformation. With the documented beta configuration, batch responses can also use the higher 300,000-token output limit.
Inputs, outputs, and reasoning
Sonnet 5.5 accepts text and images and returns text. Its image capability is for understanding visual content rather than generating images. For example, an application can use it to interpret a chart, inspect a screenshot, analyze a scanned document, or discuss an image alongside written instructions.
The model supports adaptive thinking. In practical terms, it can allocate more or less internal reasoning effort according to the task and the configured effort level. This is useful when a simple request should receive a quick answer but a difficult coding, planning, or analysis task needs more deliberate processing. Thinking behavior is not identical to ordinary visible prose: applications migrating from earlier Claude versions should review how thinking blocks are preserved, displayed, and handled between tool calls.
Anthropic’s documentation identifies structured-output workflows as supported. The supplied research does not establish a separate, general-purpose JSON mode, so developers should distinguish structured outputs from a dedicated JSON-mode setting when designing an integration.
Coding, tools, and workflow support
Coding is one of Sonnet 5.5’s main intended uses. Anthropic describes it as suitable for agentic coding, bug fixing, code review, implementation work, and maintaining codebases over longer tasks. The large context window can help an application provide more files, documentation, or repository history in a single interaction, while tool support lets the model participate in workflows that require actions outside text generation.
The model supports server-side and client-side tools, streaming, prompt caching, the Files API, PDF-oriented workflows, and batch processing. Tool use does not mean that the model independently has unrestricted access to a computer or external systems. The application still needs to define available tools, permissions, validation, and error handling.
Computer-use integrations require particular attention. On the Claude API and Google Cloud, Sonnet 5.5 supports the newer computer_toolset_20260801 toolset. The older computer_20251124 tool is rejected on those platforms. This is a concrete migration issue for applications that previously used an earlier Claude model or computer-use interface.
Sonnet 5.5 also rejects non-default values for temperature, top_p, and top_k. Applications that attempt to carry over tuning parameters from another model may therefore receive errors rather than merely different output. Developers should check the model-specific migration documentation before changing production traffic.
Strengths and trade-offs
The main strength of Claude Sonnet 5.5 is its balance. It is designed to be faster and cheaper than Claude Opus 5.5 while offering more capability than a lightweight model is intended to provide. That balance is particularly relevant for coding agents and business workflows where every request does not justify the highest available reasoning tier.
- Long-context work: The 1 million-token context window can accommodate large document sets, extensive code context, and long conversations.
- Software development: Coding, debugging, code review, and codebase maintenance are central use cases.
- Document workflows: The model can support writing, document creation, spreadsheet and presentation tasks, PDF processing, and file-based analysis.
- Visual understanding: Text-and-image input enables analysis of charts, screenshots, and other visual material.
- Agentic applications: Tools, streaming, adaptive thinking, and batch processing support applications that go beyond one-off chat responses.
- Operational cost: Its listed input and output prices are below the cost positioning expected of Anthropic’s Opus tier, while batch discounts and caching can reduce some workload costs further.
These advantages involve trade-offs. A 1 million-token context does not guarantee that every detail in a very large prompt will receive equal attention, and long outputs can still be expensive at the listed output rate. Adaptive thinking can also introduce different response behavior and integration requirements than older fixed-thinking configurations.
Limitations and compatibility considerations
Claude Sonnet 5.5 produces text only. It does not natively generate images, audio, video, music, or embeddings. Applications that require those output types need a separate model or service. Its image support should therefore be understood as visual input and image understanding, not image creation.
Sonnet 5.5 is not necessarily the best choice for every difficult task. Anthropic positions Claude Opus 5.5 above it for the most demanding reasoning and judgment-heavy work. Conversely, a lower-cost Haiku model may be more appropriate when response quality requirements are modest and request volume or latency dominates the decision.
Migration from Claude Sonnet 5 requires more than changing the model identifier. Developers should review adaptive-thinking defaults, the minimum thinking setting, preserved thinking blocks, computer-use tool types, advisor compatibility, and the restrictions on sampling parameters. If an application displays text between tool calls, it should also account for the possibility that this text is returned inside thinking blocks and configure its display behavior accordingly.
When to choose Claude Sonnet 5.5
Choose Claude Sonnet 5.5 when a project needs strong coding and knowledge-work performance but also needs fast responses and a lower operating cost than a top-tier reasoning model. It is a good fit for coding agents, code review systems, long-context document analysis, professional writing tools, image-aware assistants, and business automation that calls external tools.
It is especially suitable when the same application must handle varied workloads: a short technical question, a large repository review, a document-generation task, or an image-and-text analysis request. Prompt caching can help when those workflows repeatedly reuse a long system prompt or reference collection, and the Batch API is useful for high-volume work that can run asynchronously.
Consider Claude Opus 5.5 instead when the task is unusually judgment-heavy or when maximum capability is more important than speed and cost. Consider a Haiku-tier option when the work is simple, highly repetitive, or cost-sensitive enough that Sonnet-level capability is unnecessary. Choose another model or service when the application must generate media, create embeddings, or produce native audio or video.
Overall, Claude Sonnet 5.5 is best understood as Anthropic’s fast middle-to-upper-tier model: more capable and context-rich than a lightweight option, but positioned below Opus for the hardest work. Its value depends on using the capabilities that distinguish it—long context, coding, adaptive thinking, visual understanding, and tools—while designing around its text-only output and model-specific integration rules.