Claude Sonnet 4.6 Arrives With 1 Million Token Context and Near-Flagship Performance

Claude Sonnet 4.6 shipped on February 17, 2026, taking the default model slot for every Free and Pro user on claude.ai and raising the mid-tier bar across coding, computer use, and long-context reasoning. The release followed Claude Sonnet 4.5 by roughly five months and landed with a 79.6% score on SWE-bench Verified, a 72.5% mark on OSWorld, and a 1 million token context window arriving in beta. Pricing held at $3 per million input tokens and $15 per million output tokens, the same as Sonnet 4.5.

A Full Upgrade Across the Board

Anthropic describes Sonnet 4.6 as a complete revision rather than a targeted improvement. In internal developer tests, users preferred it over Sonnet 4.5 roughly 70% of the time on real-world coding tasks. More striking was the comparison to the previous flagship: developers chose Sonnet 4.6 over Opus 4.5 about 59% of the time in the same tests. That preference margin was large enough to shift the calculus for teams that had been defaulting to Opus on the assumption that a higher price guaranteed better results on complex tasks.

The 1 million token context window is the most structurally significant addition. At that scale, a model can hold an entire large codebase, a complete case file alongside the relevant statutory text, or months of documentation in a single session without truncation. Anthropic paired it with an automated context compaction protocol that manages the token budget as a conversation grows, preventing the performance degradation older models showed when approaching their limits. Safety evaluations showed Sonnet 4.6 outperforming Sonnet 4.5 on resistance to prompt injection attacks while matching Opus 4.6 overall.

Computer Use and Agentic Workflows

The OSWorld score of 72.5% represents a meaningful jump in computer-use capability, the ability to navigate real operating-system interfaces, fill out multi-step web forms, switch between applications, and chain browser actions into a coherent workflow. Users working with the model on computer-use tasks report that the practical bottleneck has moved from the model's ability to understand the interface to access controls and tooling, suggesting the model has cleared the comprehension threshold for most standard desktop and browser tasks.

For teams running Claude inside agentic frameworks, the upgrade path from Sonnet 4.5 was a single identifier change. Teams using Claude Managed Agents found Sonnet 4.6 reduced the number of tool calls needed to complete common multi-step tasks, cutting latency and inference cost even before the context compaction protocol is applied. Domain experts across finance, law, medicine, and engineering rated its knowledge and reasoning substantially higher than Sonnet 4.5 on sector-specific benchmarks.

"Developers preferred Sonnet 4.6 over Opus 4.5 about 59% of the time in real-world coding tests — a margin large enough to make it the default choice for most workloads." Anthropic, Claude Sonnet 4.6 launch announcement, February 2026

Pricing and Where Sonnet Fits in the Lineup

Holding pricing flat while upgrading capabilities is a pattern Anthropic has repeated across its model family, and Sonnet 4.6 continues it. At $3 per million input tokens and $15 per million output tokens, with up to 90% savings through prompt caching and 50% through batch processing, Sonnet 4.6 sits at roughly one-fifth the cost of Opus 4.6 while delivering near-parity on most practical workloads. For enterprise teams running high-volume inference, the unit economics make Sonnet 4.6 the default for everything except tasks that specifically benefit from the headroom the flagship provides.

Making Sonnet 4.6 the default for Free and Pro users means the improvement was immediately visible to the largest share of Claude's user base, not just developers with API access. It also signals Anthropic's intent to keep the mid-tier model genuinely competitive with the flagship rather than treating it as a placeholder between Opus releases. Claude Opus 4.7 arrived three months later in May 2026 with an 87.6% SWE-bench score and 3x higher image resolution at the same price, continuing the pace Sonnet 4.6 helped establish. Together, the two releases suggest Anthropic intends the Sonnet tier to carry a meaningful share of real workloads rather than serving only as a cost-reduction fallback for simpler tasks.