Anthropic releases Claude 4.5 Sonnet and the working frontier shifts again

On 29 September 2025, Anthropic released Claude 4.5 Sonnet. The naming was deliberately understated. The capability gain was not. On SWE-bench Verified, Claude 4.5 Sonnet scored 77.2 per cent, up from 72.7 per cent for Claude 4 Sonnet four months earlier. The same model was the first to clear OSWorld's standard agentic-task benchmark above the fifty per cent line.

The release also expanded Claude.ai's developer-tooling story. The Claude Code agent, launched alongside Claude 4 in May, was upgraded to use the new model by default and gained an extended-thinking mode for long-running refactors. Claude Code's monthly active developer count had reached, according to a same-day Anthropic blog post, more than half a million by late September 2025.

Computer Use, materially better

The most operationally consequential improvement was on Computer Use, the agentic desktop-control feature that Anthropic had launched in October 2024. Claude 4.5 Sonnet's score on OSWorld, the standard benchmark for AI agents performing desktop-software tasks, jumped from the high-thirties Claude 4 had managed to 61.4 per cent. As Wired and The Information observed in coverage that week, the practical effect was that agentic-AI workflows that had required a human in the loop every few minutes could, on Claude 4.5, run for thirty-to-forty minutes unattended on routine office tasks.

The benchmark name says incremental. The reliability improvement says generation.

Anthropic's enterprise revenue, by late September 2025, had crossed an annualised seven billion US dollars per The Information's reporting. Claude 4.5 Sonnet's release coincided with the run-up to the company's Series F mega-round, which closed at a one-hundred-and-eighty-three-billion-dollar post-money five days before this release. The two events are inseparable in any narrative of late-2025 frontier AI economics.