September's biggest release came at the very end. On the 29th, Anthropic launched Claude Sonnet 4.5 and called it the best coding model in the world.¹
The Benchmarks
Anthropic says Sonnet 4.5 is state of the art on SWE-bench Verified, the standard test for fixing real GitHub issues. On OSWorld, which tests models on real computer tasks, it scored 61.4%. Four months ago, Sonnet 4 led that benchmark at 42.2%.
OSWorld matters more to me than the coding score. Most real work happens in messy software, and OSWorld is closer to that than a clean test harness.
⏰ 30 Hours on One Task
Anthropic says it has seen Sonnet 4.5 stay focused on complex, multi-step tasks for more than 30 hours. In early trials with enterprise customers, it built an application, stood up database services, bought domain names, and ran a SOC 2 audit on its own.²
That changes how you work with the model. It acts less like a pair programmer you chat with and more like a contractor you hand a specification to. The hard part moves to writing a good spec and checking the result.
The Claude Agent SDK
Anthropic also released the Claude Agent SDK, the same infrastructure that powers Claude Code.¹ Teams can build their own agents on it, for coding or anything else.
Pricing and Availability
Pricing stayed at $3 per million input tokens and $15 per million output tokens, the same as Sonnet 4. It launched in the Claude apps and API and in Amazon Bedrock and Google Cloud's Vertex AI. Microsoft added it to Copilot Studio the same day.³
Where Anthropic Stands
Early in the month, Anthropic raised $13 billion at a $183 billion valuation and said its run-rate revenue had passed $5 billion.⁴ Menlo Ventures' mid-year survey put Anthropic at 42% of enterprise code generation, twice OpenAI's share.⁵ Sonnet 4.5 also lands a week before OpenAI's DevDay.
What It Means for Software Teams
If a model can work for a day on its own, the bottleneck moves from writing code to deciding what to build and checking what got built. For product managers, that makes specs, acceptance criteria, and review more important. Testing an idea with a working prototype in a day could also change how product discovery works.
What would you hand an agent for 30 hours?