Anthropic unveiled Claude Sonnet 4.5 on September 29, 2025. The company says the new model is its strongest model for coding and for building long-running agents. Anthropic adds that Sonnet 4.5 shows big gains on real-world coding tests and on tasks that need sustained focus and tool use.
Claude Sonnet 4.5 is available today through the Claude API and in the Claude apps. The developer price remains the same as the previous Sonnet release. Anthropic lists the API name as claude-sonnet-4-5 and the token price as three dollars per million input tokens and fifteen dollars per million output tokens.
New coding model
Anthropic reports strong benchmark results. The company says Sonnet 4.5 leads on the SWE Bench Verified evaluation for software engineering tasks. The firm also documents higher scores on OSWorld and other measures that test real computer use. Anthropic states that the model can stay on task for many hours and that it handled more than thirty hours of autonomous coding in early trials.
Anthropic shares customer feedback that highlights longer-horizon work and better tool use. Partners told Anthropic that Sonnet 4.5 can manage complex multi-step problems. The firm also says the model is more efficient when it uses external tools like terminals and file editors inside a conversation. These features aim to let the model act more like a software engineer who can run commands and edit files.
Agent SDK launch
Along with Sonnet 4.5, Anthropic released the Claude Agent SDK. The SDK is the same infrastructure that powers Claude Code. Anthropic says the SDK helps developers build agents that keep memory over long sessions and that coordinate subagents working on shared goals. The company positions the SDK as a building block for agentic apps beyond coding.
Anthropic also put a short research preview in the wild called Imagine with Claude. The preview shows Sonnet 4.5 creating software on the fly. The company made the preview available to Max subscribers for a limited time to demonstrate agentic creation and real-time code generation.
Safer and more aligned
Anthropic says Sonnet 4.5 is its most aligned frontier model to date. The company reports reduced rates of sycophancy and deception and improved defenses against prompt injection attacks. Anthropic notes that Sonnet 4.5 ships under its AI Safety Level 3 protections and comes with classifiers aimed at high-risk areas. The firm also published a detailed system card and evaluations for transparency.
What this means for developers
Developers who build coding tools and agents may find Sonnet 4.5 useful for long workflows and for tasks that need reliable tool use. Anthropic positions the model as a drop-in upgrade for existing Sonnet users. The firm says code execution, file creation, and the new agent capabilities are available across paid plans and in the Claude apps and API.
Anthropic launched Sonnet 4.5 quickly after its prior model. The fast cadence highlights the competitive pace of the AI market. For teams that rely on code agents, this release signals new options for production-grade automation and for building agentic developer tools.
