Anthropic launches Claude Sonnet 4.5, leading SWE-bench Verified and OSWorld
Claude Sonnet 4.5 posted the highest results on SWE-bench Verified and OSWorld at launch, with gains for long coding sessions and computer use.
Anthropic has announced Claude Sonnet 4.5. At the time of launch, the model leads two evaluations, SWE-bench Verified and OSWorld. According to the company, Sonnet 4.5 is also improved for long-running coding sessions and for computer use capability.
SWE-bench Verified requires a model to fix bugs in real software projects, so a high score implies the ability to read and understand source code and produce a correct patch. OSWorld measures the ability to control a computer through its interface, meaning the model operates the screen and applications itself rather than only generating text.
For businesses, these are two categories of work that consume a lot of staff time: source code maintenance, and repetitive data entry and operations in software. If a model can handle a long sequence of operations without drifting off course, engineering teams can move part of the mechanical work to the tool.
What to keep in mind is that results on a benchmark cannot replace testing in a real environment. For computer-use tasks, the scope of permissions and the stopping points should be defined in advance, together with a human review step.
Source: anthropic.com
