Claude Sonnet 4.6

At release, the most capable Sonnet model Anthropic had shipped, upgrading coding, computer use, long-context reasoning, agent planning, knowledge work and design in one step. Its consistency and instruction-following gains led early-access developers to prefer it to its predecessor by a wide margin, and often to Opus 4.5. Thinking is off by default; effort runs low, medium, high and max, with no xhigh rung on this model.

Sonnet 4.6 was the balanced middle of Anthropic's line from February 2026 until Sonnet 5 replaced it at the end of June, and it is still in the catalogue as the previous Sonnet generation. It was the release that brought a 1M token context window to the Sonnet tier in beta, and Anthropic made it the default on the Free and Pro plans on the day it shipped.

Computer use is the capability the post is built around. Anthropic's argument is that most organisations have specialised software predating modern interfaces, which previously needed a bespoke connector for each system; a model that clicks and types the way a person does removes that step. The lab reports 72.5% on OSWorld-Verified against 61.4% for Sonnet 4.5, and describes early users seeing human-level work on tasks like navigating a complex spreadsheet or completing a multi-step web form and then pulling the result together across several browser tabs. It says plainly that the model still lags the most skilled humans, and that OSWorld measures a controlled set of tasks while real computer use is messier, more ambiguous and carries higher stakes for errors.

Preference data is the other half of the case. In Anthropic's early Claude Code testing, users chose Sonnet 4.6 over Sonnet 4.5 roughly 70% of the time and over Opus 4.5 - the frontier model from three months earlier - 59% of the time, reporting less overengineering, better instruction following, fewer false claims of success and more consistent follow-through on multi-step work. On the published table the lab measures 79.6% on SWE-bench Verified, 89.9% on GPQA Diamond, 1633 on GDPval-AA Elo and 63.3% on Finance Agent v1.1, the last two the best figures in its comparison set.

Two results say something about the model's character. On Vending-Bench Arena, which pits models against each other running a simulated business, Anthropic reports Sonnet 4.6 inventing a strategy of its own: spending heavily on capacity for the first ten simulated months, well beyond what its competitors spent, then pivoting sharply to profit in the final stretch, with the timing of the pivot carrying it clear of the field. And on safety, the lab reports a major improvement over Sonnet 4.5 in resisting prompt injection, putting it near Opus 4.6 - which matters precisely because computer use is the thing it is best at.

Anthropic says Sonnet 4.6 performs strongly at any thinking effort, extended thinking off included, and advised anyone migrating from Sonnet 4.5 to try the whole spectrum rather than assume a setting. That is worth knowing here, where thinking is off unless a call asks for it and the ladder runs low, medium, high and max with no extra rung.