Claude Sonnet 4.5 Released: The Coding Model That Works Autonomously for 30 Hours

Anthropic released Claude Sonnet 4.5 on September 29, 2025, calling it the best coding model in the world at the time: it topped SWE-bench Verified and could work autonomously on complex tasks for more than 30 hours — long-horizon autonomy became the new competitive axis.

Anthropic simultaneously upgraded Claude Code (checkpoints, subagents) and shipped a browser extension; Sonnet 4.5 delivered flagship coding at mid-tier pricing and quickly became the default choice inside Cursor, Windsurf and other tools.

From Sonnet 4.5 (September) to Haiku 4.5 (October) to Opus 4.5 (November), Anthropic shipped three models in three months — nailing the 'best at coding' mindshare firmly to its own name.

What 30 Hours of Autonomy Actually Measures

'Works autonomously for 30 hours' is not just a duration stat — it marks a switch in the competitive axis. It measures engineering reliability across long task chains: staying on course, retaining context, and rolling back to retry when things break. That is the real bottleneck of productizing agents (see our Devin 2.0 coverage): anyone can survive a five-minute demo; products are won in hour twenty. The companion Claude Code features — checkpoints and subagents — are essentially save-state, division-of-labor and rollback scaffolding for long-horizon work.

The mid-tier pricing is equally deliberate: flagship-grade coding at Sonnet prices means Cursor, Windsurf and peers can make it the default model without destroying subscription margins (see our Windsurf acquisition coverage) — a move that bolted Claude into the coding-tool supply chain as the default option.

Our Take

Sonnet 4.5 marks the narrative switch of late 2025: beyond benchmark scores, 'how long can it work unattended' became the new hard metric — the one that decides whether agents graduate from demo to production (see our AI coding tools war coverage). It also set up Opus 4.5's discounted crown two months later (see our coverage): first occupy the toolchain default slot with a mid-tier model, then harvest mindshare with a flagship price war. Anthropic's one-two punch is the clearest template of model-vendor strategy in 2025.