What Claude Sonnet 4.6 is
Claude Sonnet 4.6 is a hybrid reasoning model provided by Anthropic. The model is intended for tasks that combine ordinary language generation with planning, analysis, coding, tool use, and multi-step execution. Its canonical Claude API model ID is claude-sonnet-4-6. The model was released on February 17, 2026.
In Anthropic’s lineup, Sonnet represents a middle position between lower-cost, faster models and the provider’s more expensive Opus-class models. That positioning makes Sonnet 4.6 a practical choice when an application needs substantial reasoning and coding capability but must also control latency or token costs. Anthropic describes improvements over Sonnet 4.5 in areas including instruction following, agent planning, computer use, document comprehension, financial analysis, and professional workflows.
There is an important lifecycle qualification. The model remains accessible as of September 24, 2026, but Anthropic classifies it as an active legacy model and recommends migration to Claude Sonnet 5. Anthropic’s documentation states that retirement is not expected sooner than February 17, 2027. The fixed model ID identifies a specific 4.6-generation snapshot rather than a rolling alias that automatically changes to a newer model.
Specifications at a glance
| Specification | Claude Sonnet 4.6 |
|---|---|
| Provider | Anthropic |
| Release date | February 17, 2026 |
| Model ID | claude-sonnet-4-6 |
| Context window | 1,000,000 tokens |
| Standard maximum output | 128,000 tokens |
| Batch maximum output | 300,000 tokens in beta |
| Input | Text and images |
| Output | Text |
| Reasoning | Adaptive thinking with high default effort |
| Reliable knowledge cutoff | August 2025 |
| Training-data cutoff | January 2026 |
| Lifecycle | Active legacy model |
The reliable knowledge cutoff and training-data cutoff are different pieces of information. Anthropic lists August 2025 as the reliable knowledge cutoff, while the training-data cutoff is January 2026. External tools such as web search can provide newer information during a task, but they do not change the model’s underlying knowledge cutoff.
Context window and output limits
Sonnet 4.6 supports a 1-million-token context window. A context window is the amount of material the model can consider within a request and its surrounding conversation. In practical terms, this capacity can support large software repositories, lengthy contracts, research collections, visual documents, and extended agent sessions without requiring the user to divide all material into small, disconnected prompts.
The standard maximum output is 128,000 tokens. This is a ceiling rather than a recommendation that every response should be that long. Long outputs can increase cost, take more time to generate, and become harder to review. Anthropic also documents a 300,000-token maximum output for Batch API beta requests, which is distinct from the standard interactive limit.
A large context window does not make every large input equally useful. Applications still need sensible document selection, retrieval, summarization, and output controls. Sending more material can also increase input-token charges, so the window is best treated as capacity for demanding workflows rather than a reason to include irrelevant content.
Pricing and usage economics
Anthropic’s standard API pricing for Sonnet 4.6 is $3 per million input tokens and $15 per million output tokens. Input tokens are the text, image-related content, instructions, documents, and conversation history supplied to the model. Output tokens are the generated response. The total cost of a request depends on both quantities, not simply on the model selected.
Prompt caching can reduce the cost of repeatedly sending the same context. Anthropic documents a price of $3.75 per million tokens for a five-minute cache write, $6 per million tokens for a one-hour cache write, and $0.30 per million tokens for cache reads. Batch API requests receive a 50% discount on input and output token pricing. These options are most relevant to applications that reuse large prompts, process workloads asynchronously, or handle many similar requests.
The pricing helps explain Sonnet 4.6’s role in a model lineup. It is more capable and more expensive than a minimal model may be, but it is positioned below Anthropic’s highest-cost Opus tier. For production systems, the relevant trade-off is not only price per million tokens. Teams should also consider how much additional prompting, retrying, tool orchestration, or human review a less capable model would require.
Reasoning, coding, and agentic work
Sonnet 4.6 uses adaptive thinking. This allows the model to allocate reasoning effort according to the task rather than treating every request identically. Anthropic describes the default effort as high, while also noting that extended thinking is deprecated for this model. In practical use, its reasoning capabilities are relevant to tasks such as decomposing a software change, weighing several possible approaches, checking dependencies, and carrying a plan through multiple tool calls.
Coding is one of the model’s strongest intended use cases. Sonnet 4.6 is suited to implementing features, debugging, refactoring, reviewing code, maintaining existing applications, and developing front-end interfaces. It can be useful when the task requires more than producing a short code fragment—for example, understanding a repository, identifying related files, changing an implementation, and explaining the consequences of the change.
Agentic workflows use a model as part of a larger system that can inspect information, call tools, and take a sequence of actions. Sonnet 4.6 supports tool use, including function-style tool calls and Anthropic’s supported tool ecosystem. Strict tool use and structured outputs can help applications receive data in a predictable schema rather than parsing an unrestricted paragraph. Structured outputs should not automatically be described as a separate legacy JSON-mode capability; the supplied research does not verify a distinct JSON mode for this model.
Vision and supported modalities
Sonnet 4.6 accepts both text and image input and produces text output. Image understanding can be used for screenshots, charts, tables, visual documents, and supported PDF content. This makes the model suitable for workflows such as examining a user-interface screenshot, extracting information from a scanned document, or combining written instructions with a diagram.
The model does not natively generate images, video, audio, speech, music, or embeddings. This limitation matters when selecting a model for a complete media pipeline. Sonnet 4.6 can describe or analyze visual material, but a separate image or audio-generation system is required when the desired result is a new image, voice recording, soundtrack, or video.
Where Sonnet 4.6 is available
Anthropic makes Sonnet 4.6 available through the Claude API and lists availability through Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. The platform-specific names and configuration details can differ, but the canonical Claude API identifier is claude-sonnet-4-6. The Amazon Bedrock identifier supplied in the research is anthropic.claude-sonnet-4-6.
Availability through a cloud platform does not necessarily mean that every feature has identical behavior in every environment. Tool support, web search, structured outputs, regional access, quotas, and lifecycle policies may depend on the platform and account configuration. Teams should verify the target platform’s current documentation before committing to a feature-specific integration.
Best use cases
- Repository-scale coding: reviewing, refactoring, debugging, and extending code across many files.
- Agentic software development: planning implementation steps, using tools, checking results, and revising a solution.
- Long-document analysis: examining contracts, research collections, policies, or technical documentation within a single extended context.
- Computer-use workflows: interpreting visual interfaces and coordinating actions through supported tools.
- Enterprise knowledge work: research, financial analysis, document comprehension, structured extraction, and professional writing.
- High-volume production applications: workloads that need a balance of reasoning quality, latency, context capacity, and Sonnet-tier pricing.
The model is particularly attractive when a task requires both analysis and execution. For example, an application might provide a large codebase, ask the model to identify a bug, call a repository tool, propose a patch, and return a structured summary for review.
Limitations and when to choose another option
Sonnet 4.6 is not the right choice for every workload. Its text-only output means that applications requiring native image, video, speech, music, or audio generation need another model or an additional service. It also should not be treated as an unrestricted source of current facts: its knowledge cutoff is limited, and even web-assisted workflows require verification for important decisions.
Usage limits can vary by plan, account, conversation size, and platform. Advanced features may be paid, beta, or restricted by location. In consumer settings, privacy and model-improvement policies also differ from commercial API arrangements, so organizations should review the applicable terms before sending sensitive information.
For a new project, Claude Sonnet 5 deserves direct evaluation because Anthropic recommends it as the successor and current Sonnet model. Sonnet 4.6 can still make sense when an existing application depends on its fixed behavior, when migration has not yet been completed, or when testing shows that its price, performance, and compatibility are a better fit. A faster or cheaper model may be preferable for simple classification, short transformations, or low-risk high-volume tasks. A higher-tier reasoning model may be preferable for unusually difficult analysis when the additional cost is justified. These are deployment trade-offs rather than claims that one model wins every task.
Overall assessment
Claude Sonnet 4.6 is best understood as a capable middle-tier model for serious coding, long-context analysis, computer use, and tool-driven automation. Its strongest practical distinctions are the 1-million-token context window, adaptive reasoning, text-and-image input, structured tool-oriented workflows, and pricing below Anthropic’s highest-end models. Those benefits are balanced by text-only output, usage and availability constraints, a finite knowledge cutoff, and its active-legacy status.
It remains a reasonable model to understand and evaluate for existing systems, especially where its fixed model ID and large context are important. For new deployments, however, the lifecycle warning is material: Anthropic recommends assessing Claude Sonnet 5 rather than assuming Sonnet 4.6 is the long-term default.