Anthropic announced Claude Sonnet 4.5 on September 29, 2025. The company positions this release as a model optimized for software development, computer use, and extended multi-step tasks, with improvements in reasoning and mathematics compared to prior models.
Availability and pricing
Claude Sonnet 4.5 is available everywhere as of September 29, 2025. Developers can access the model via the Claude API using the claude-sonnet-4-5 identifier. Pricing is unchanged from Claude Sonnet 4: $3/$15 per million tokens.
Product updates released alongside the model
Anthropic shipped multiple product upgrades together with Sonnet 4.5:
- Claude Code: added checkpoints (a highly requested feature), a refreshed terminal interface, and a native VS Code extension.
- Claude API: new context editing and memory tools to let agents run longer and manage greater complexity.
- Claude apps: code execution and file creation (spreadsheets, slides, documents) available directly in the conversation UI.
- Claude for Chrome extension: rolled out to Max users who joined the waitlist.
- Claude Agent SDK: the infrastructure behind Claude Code is now available as an SDK for developers to build agents.
Performance: coding, computer tasks, and reasoning
Anthropic reports strong benchmark results for Sonnet 4.5. On the SWE-bench Verified evaluation (a 500-problem dataset for real-world software tasks) they report a 77.2% average score using a 200K thinking-token budget, averaged over 10 trials with a simple tool scaffold (bash and string-based file edits). A 1M context configuration achieves 78.2%, and a “high compute” evaluation that uses multiple parallel attempts and an internal scoring model reaches 82.0%.
On OSWorld, which measures real-world computer task performance, Sonnet 4.5 scores 61.4%, up from Sonnet 4’s 42.2 four months earlier.
The company states the model can maintain focus for more than 30 hours on complex, multi-step tasks and shows improvements in reasoning and math across a broad suite of evaluations.
Customer and expert feedback
Anthropic published several customer and expert testimonials from finance, law, medicine and STEM fields, describing improved domain knowledge and reasoning versus older models such as Opus 4.1. Examples included:
- Improved multi-step reasoning and code comprehension for GitHub Copilot integrations (Mario Rodriguez, Chief Product Officer).
- A 44% reduction in average vulnerability intake time and a 25% accuracy improvement for Hai security agents (Nidhi Aggarwal, Chief Product Officer).
- Internal code-editing benchmark error rate reduced from 9% on Sonnet 4 to 0% on Sonnet 4.5 (Michele Catasta, President).
- The ability to sustain 30+ hours of autonomous coding, enabling engineers to compress months of architectural work while maintaining coherence across large codebases (Sean Ward, CEO and Co-Founder).
Other customers reported gains such as an 18% improvement in planning performance and a 12% increase in end-to-end evaluation scores for a planning agent, along with early improvements in Figma and Canva product integrations.
Alignment and safety
Anthropic describes Sonnet 4.5 as their most aligned frontier model to date. They report reductions in concerning behaviors including sycophancy, deception, power-seeking, and encouragement of delusional thinking, and note progress on defenses against prompt-injection attacks, a major risk for agentic and computer-use capabilities.
The model is released under Anthropic’s AI Safety Level 3 (ASL-3) protections. These include classifier-based filters intended to detect potentially dangerous inputs and outputs, particularly those related to chemical, biological, radiological, and nuclear (CBRN) hazards. Anthropic says it has reduced false positives in these classifiers by a factor of ten compared to their original description and by a factor of two since Claude Opus 4 was released in May 2025. For cases where classifiers interrupt interactions, users can continue conversations with Sonnet 4, a model with lower CBRN risk.
A detailed set of safety and alignment evaluations, including tests using mechanistic interpretability techniques, is available in the Claude Sonnet 4.5 system card.
Claude Agent SDK and research preview
Anthropic is releasing the Claude Agent SDK — the same infrastructure that powers Claude Code — to let developers build agents that manage memory across long-running tasks, balance autonomy and permissions, and coordinate subagents toward shared goals.
Alongside Sonnet 4.5, Anthropic launched a temporary research preview called "Imagine with Claude." In this experiment the model generates software in real time without prewritten code. The preview is available to Max subscribers for five days at claude.ai/imagine.
Recommendation and technical documentation
Anthropic recommends upgrading to Sonnet 4.5 across uses, since it acts as a drop-in replacement with improved performance at the same price. Claude Code updates are available to all users; developer platform updates and the Claude Agent SDK are available to all developers. Code execution and file creation in the Claude apps are included on all paid plans.
For complete technical details and evaluation results, Anthropic points readers to the Sonnet 4.5 system card, model page, documentation, engineering posts, and related research on cybersecurity.
Methodology notes (selected)
- SWE-bench Verified: reported 77.2% with the 200K thinking budget averaged over 10 trials on the full 500-problem dataset; the 1M context configuration achieves 78.2%. The high-compute procedure described uses parallel sampling, rejection of patches that fail visible tests, and internal scoring to select the best candidate (82.0%).
- OSWorld: scores use the official OSWorld-Verified framework with 100 max steps, averaged across 4 runs.
- Other benchmarks mentioned include Terminal-Bench, τ2-bench, AIME, MMMLU and a Finance Agent leaderboard run by Vals AI; Anthropic provides additional configuration details and caveats in its documentation.
Summary
Claude Sonnet 4.5 is presented by Anthropic as a major incremental step for coding, long-context reasoning and agent workflows, released with a suite of product upgrades (checkpoints, Agent SDK, Chrome extension and a short research preview). The model is available now at the same token pricing as the previous generation and includes detailed safety controls consistent with Anthropic’s ASL-3 framework.
