What is Claude Sonnet 4.5?
Claude Sonnet 4.5 is Anthropic’s Claude 4.5 Sonnet model, released on September 29, 2025. Anthropic positioned it for coding, complex agents, computer use, reasoning, and long-horizon tasks. In practical terms, it is intended for work that requires more than a single short answer: maintaining context across a large repository, planning and executing several development steps, reviewing visual material, or coordinating actions through external tools.
The model belongs to Anthropic’s Claude family and is accessed primarily through the Claude API and supported cloud platforms. It is not a standalone image, audio, or video generator. Its multimodal capability refers to its ability to accept both text and images while returning text.
Anthropic currently classifies Claude Sonnet 4.5 as a legacy model and recommends migration to Claude Sonnet 5. It remains available according to the supplied lifecycle information, with retirement scheduled no sooner than September 29, 2026; Anthropic has not published an exact shutdown date.
Key specifications at a glance
| Specification | Claude Sonnet 4.5 |
|---|---|
| Provider | Anthropic |
| Release date | September 29, 2025 |
| Canonical API model ID | claude-sonnet-4-5-20250929 |
| Context window | 200,000 tokens |
| Maximum output | 64,000 tokens |
| Input | Text and images |
| Output | Text, including structured responses |
| Reasoning | Extended thinking supported |
| Status | Legacy, but still available according to the supplied research |
A token is a unit of text used for model processing; it may represent a word, part of a word, punctuation, or other text fragment. The 200,000-token context window covers the material the model can consider in a request and conversation, while the 64,000-token maximum applies to the response it can produce in one request.
Why it is suited to coding and agent work
Claude Sonnet 4.5’s main distinction is its focus on coding and long-running agentic workflows. A coding agent is an application that combines a language model with tools such as a file system, terminal, browser, issue tracker, or source-control workflow. Sonnet 4.5 can help inspect a repository, identify relevant files, propose a plan, write or modify code, interpret test output, and continue through several related steps.
Its large context window is useful when a task involves extensive source code, technical documentation, logs, or design material. The model can also analyze screenshots, diagrams, interfaces, and visual PDF content when those are supplied as supported image input. This makes it suitable for tasks such as explaining an interface screenshot, reviewing a visual error report, or combining a written specification with a diagram.
Tool use does not mean that the base model independently operates a computer or gains unrestricted access to external systems. An application must provide client-side tools or enable compatible Anthropic server tools. Developers remain responsible for permissions, authentication, sandboxing, confirmation steps, and validation of any action that could modify data or affect an external system.
Reasoning, tools, and structured output
Claude Sonnet 4.5 supports extended thinking, which allows the model to spend additional processing effort on difficult problems before producing its answer. This is useful for multi-step coding tasks, debugging, planning, research synthesis, and problems where a quick response is more likely to miss dependencies or constraints. Extended thinking can increase the amount of processing required, so it should be enabled selectively when the additional reasoning is worth the cost or latency.
The model supports client-side and server-side tool use, streaming responses, structured outputs, strict tool use, prompt caching, and Anthropic’s Message Batches API. Structured outputs are useful when an application needs predictable fields rather than free-form prose, such as extracting an issue list, returning a test plan, or classifying documents. The supplied research does not verify a separate legacy “JSON mode” capability, so structured outputs should not automatically be treated as the same feature.
Prompt caching can reduce repeated processing for stable instructions or large prefixes that are reused across requests. The Message Batches API offers a 50% discount on input and output token pricing, according to the supplied pricing information, but batch processing is intended for workloads that do not require an immediate individual response.
Supported modalities and practical limits
- Text input: Supported.
- Image input: Supported for visual analysis alongside text.
- Text output: Supported, including code and structured responses.
- Image, audio, and video output: Not natively supported.
- Context: Up to 200,000 tokens.
- Maximum response: Up to 64,000 output tokens.
These limits make Sonnet 4.5 appropriate for substantial documents, codebases, and multi-step conversations, but they do not guarantee that every task will fit comfortably in one request. Large inputs still need careful organization, and application developers must account for tool results, instructions, conversation history, and the requested response within the available context.
API pricing
Anthropic lists Claude Sonnet 4.5 at $3 per million input tokens and $15 per million output tokens. Input and output are priced separately, so an application that generates long responses can spend considerably more on output than on the corresponding input.
| Usage type | Price |
|---|---|
| Input tokens | $3 per million tokens |
| Output tokens | $15 per million tokens |
| Five-minute prompt-cache write | $3.75 per million tokens |
| One-hour prompt-cache write | $6 per million tokens |
| Prompt-cache read | $0.30 per million tokens |
| Message Batches API | 50% discount on input and output token pricing |
Prompt caching is most relevant when the same large instructions or reference material are sent repeatedly. The cache-write rates are higher than the normal input rate, while cache reads are substantially cheaper. The right choice therefore depends on how often the cached material is reused and how long it remains valid.
Strengths and trade-offs
The supplied research rates Sonnet 4.5 highly for coding and complex reasoning, but those scores are editorial evaluations rather than Anthropic-published specifications. The concrete strengths supported by the documentation are its 200,000-token context, 64,000-token output limit, image understanding, extended thinking, tool support, structured outputs, prompt caching, and batch processing.
Its principal trade-off is cost and latency relative to smaller or simpler models. The model is designed for difficult tasks rather than the lowest-cost response to every request. Extended thinking, large tool traces, and long generated outputs can increase both processing time and token usage. Applications that only need short classification, simple extraction, or brief routine replies may not benefit from using a model aimed at complex agent workflows.
Sonnet 4.5 also cannot replace a native media-generation model. It can analyze images, but it does not natively create images, audio, or video. Its context window is large but does not exceed 200,000 tokens, so tasks requiring a larger working context may need a different option or a retrieval and summarization strategy.
Best use cases
- Repository-scale coding, refactoring, debugging, and code review.
- Software agents that need to inspect files, run tools, and maintain a multi-step plan.
- Computer-use or browser workflows where the application controls permissions and actions.
- Technical research and document analysis involving substantial written material.
- Visual analysis of screenshots, diagrams, interfaces, and supported PDF content.
- Applications requiring structured responses or strict tool invocation.
- Long-running professional workflows where consistency across many steps matters.
For example, a development assistant could receive a feature request, inspect relevant files through application-provided tools, propose an implementation plan, modify code, interpret test failures, and return a structured summary of the changes. Human review and execution safeguards remain important, especially when tools can change production systems or access sensitive data.
When to choose Claude Sonnet 4.5
Choose Claude Sonnet 4.5 when the workload benefits from strong coding and reasoning performance, image understanding, a large context window, tool use, and sustained multi-step interaction. It is a reasonable fit for teams that need a capable general model for software agents and complex analysis and that can accept its higher output price compared with lighter models.
Consider another option when the primary requirement is the lowest latency or lowest cost, when the task needs native image, audio, or video generation, or when a context window larger than 200,000 tokens is essential. New deployments should also evaluate Claude Sonnet 5 because Anthropic recommends migration to that newer Sonnet model. Sonnet 4.5 may still be appropriate where an existing application depends on its behavior, pricing, integrations, or validated workflow, but its legacy status makes lifecycle planning important.
Availability and model identity
The dated API identifier for the model is claude-sonnet-4-5-20250929. Anthropic also documents claude-sonnet-4-5 as a convenience alias that points to the most recent dated snapshot for this minor version. Using the dated identifier can make deployments more reproducible, while an alias may follow Anthropic’s designated snapshot for the model family.
Sonnet 4.5 is available through Anthropic’s developer platform and partner cloud services according to the supplied research. Because it is classified as legacy and has a retirement window beginning no sooner than September 29, 2026, developers should check Anthropic’s lifecycle documentation before committing it to a new long-lived system.