On September 29, 2025, Anthropic released Claude Sonnet 4.5 and described it as the best coding model in the world, its strongest model for building complex agents and its best model at using computers. The notable part was the price tier: Sonnet 4.5 kept the $3 and $15 per million token pricing of Claude Sonnet 4 while, by Anthropic's own measurements, outperforming the more expensive Claude Opus 4.1 in several domain evaluations. Anthropic paired the model with a wave of developer tooling, including the Claude Agent SDK, which made the launch as much about building agents as about the model itself.
Key Facts#
- Release date: September 29, 2025. API alias
claude-sonnet-4-5, with the dated IDclaude-sonnet-4-5-20250929. - Pricing per million tokens: $3 input and $15 output, the same as Claude Sonnet 4.
- Vendor-reported coding results: 77.2% on SWE-bench Verified, or 82.0% with parallel test-time compute.
- Computer use: 61.4% on the OSWorld benchmark, up from 42.2% for Claude Sonnet 4 four months earlier.
- Long tasks: Anthropic says it observed the model staying focused for more than 30 hours on complex, multistep tasks.
- Safety level: released under Anthropic's AI Safety Level 3 protections, which include classifiers for chemical, biological, radiological and nuclear risks.
- Tooling announced alongside: the Claude Agent SDK, context editing and a memory tool in the API, checkpoints and a native VS Code extension for Claude Code, file creation in the Claude apps, and a Claude for Chrome extension for Max users.
What Happened#
Anthropic framed Sonnet 4.5 as a step change for agents rather than chat. Beyond the SWE-bench Verified result, the company highlighted OSWorld, a benchmark that tests models on real-world computer tasks. Sonnet 4.5 scored 61.4%, a gain of more than 19 points over Claude Sonnet 4 in four months. It also said Sonnet 4.5 improved on Claude Opus 4.1 in domain-specific evaluations for finance, law, medicine and STEM work.
The 30-hour figure drew attention because it spoke to endurance. Coding agents often fail not on a single hard step but on drift: losing track of goals, repeating work or giving up. Anthropic said it had seen Sonnet 4.5 maintain focus for more than 30 hours on complex tasks, although that is an observation from Anthropic's own testing rather than a standardized benchmark.
Anthropic also called Sonnet 4.5 the most aligned frontier model it had released. It reported reductions in sycophancy, deception, power-seeking and the tendency to encourage delusional thinking, along with progress in defending against prompt injection, a critical property for any agent that reads web pages, emails or documents. The company said its safety classifiers now produced ten times fewer false positives than when first described, and half as many as at the Claude Opus 4 launch.
The tooling mattered as much as the model. The Claude Agent SDK packaged the infrastructure behind Claude Code so developers could build their own agents on it. In the API, context editing and a memory tool addressed a practical problem in long-running agents: managing what stays in the context window as a task grows. Claude Code gained checkpoints, which let users roll back to an earlier state, and a native VS Code extension. Launch customers quoted in the announcement included Cursor, GitHub, Replit, Canva and Figma.
Background#
Sonnet 4.5 arrived four months after Claude Opus 4 and Sonnet 4 and less than two months after Claude Opus 4.1. In the meantime, OpenAI had launched GPT-5 in August with strong coding results and aggressive pricing, raising the competitive bar for every coding assistant built on Anthropic models.
The launch showed Anthropic pushing its mid-priced Sonnet line hard. Because most coding tools and agent products run huge volumes of requests, a Sonnet-tier model that matches or beats Opus-tier results changes the economics of the entire category.
Why It Matters for Developers#
The clearest lesson from Sonnet 4.5 is that the price tier no longer predicts capability. Teams that had defaulted to the most expensive model for hard tasks needed to re-run their evaluations, because a cheaper model could now win. Build that into your process: keep a benchmark set drawn from your own backlog and re-test whenever a new model appears.
For .NET developers building agents, the launch highlighted three design problems that any framework must handle: long-running state, context growth and tool safety. Anthropic's context editing and memory tool are one answer. On the .NET side, the Microsoft Agent Framework offers its own abstractions for agent threads, tools and workflows, and the official Anthropic C# SDK implements IChatClient, so Claude models can sit behind the same abstraction as other providers. Our guide to AI agent architecture patterns explains how to decide what the model should remember and what your application should store.
The computer-use results are relevant for enterprises with legacy desktop software that has no API. An agent that can operate a UI reliably opens automation options, but it also widens the attack surface: prompt injection through on-screen content is a real risk. Apply least privilege, sandbox the environment and log every action, as described in our guide to responsible AI and LLM security.
Finally, the Claude Code updates, checkpoints in particular, reflect a broader shift toward reviewable, reversible agent work. If your team is adopting coding agents, our guide to AI-assisted .NET development covers review workflows and guardrails.
What's Next#
Anthropic followed Sonnet 4.5 with Claude Haiku 4.5 in October 2025 and Claude Opus 4.5 in November 2025, and later released Claude Sonnet 4.6 in February 2026 and Claude Sonnet 5 in June 2026. Sonnet 4.5 itself remains active on the Claude API, but Anthropic's model deprecation page lists its tentative retirement date as not sooner than September 29, 2026, one year after launch.
That date is a signal to plan ahead. Anthropic commits to at least 60 days' notice before retiring publicly released models, and it recommends testing replacements well before retirement. If your applications pin claude-sonnet-4-5-20250929, now is the time to evaluate a newer Sonnet model against your own test suite.