Anthropic has rolled out Claude Sonnet 4.6, the latest update to its mid-sized Sonnet model. In its announcement, the company highlights sharper coding performance, stronger instruction-following, and improved computer-use capabilities.
Anthropic says Sonnet 4.6 is now the default model on the Free tier, and it’s positioned as the go-to Sonnet option across its consumer and developer surfaces—especially for day-to-day coding and agentic workflows.
👉 Clean Code by Robert C. Martin (Amazon search)
(Affiliate link: using this link may earn a commission.)
A 1M-token context window (in beta)
The headline technical upgrade is a 1 million–token context window, available in beta. Anthropic frames that capacity as large enough to handle entire codebases, long contracts, or dozens of research papers in a single request—use cases that quickly overwhelm the typical 128k–200k windows most developers plan around.
That said, “1M tokens” in practice will still come down to cost, latency, and how reliably the model can retrieve and reason over information deep in the context. Big windows are powerful, but they also raise the bar for prompting and evaluation—especially when users expect the model to “remember” everything equally well.
Coming shortly after Opus 4.6
The Sonnet update lands just 12 days after Anthropic announced Opus 4.6 (Opus on Feb. 5, 2026, Sonnet on Feb. 17, 2026), underscoring how fast the company is iterating across its lineup. That naturally invites speculation that a Haiku refresh could be next—but Anthropic hasn’t confirmed a timetable.
Benchmarks: strong gains, with important footnotes
Anthropic is also shipping Sonnet 4.6 with a new set of benchmark results, including OSWorld-Verified (computer use) and SWE-bench Verified (software engineering). The company frames the results as meaningful progress for Sonnet’s real-world agentic performance—particularly the kind of multi-step tasks that mix reading, clicking, coding, and tool use.
The splashiest number may be 60.4% on ARC-AGI-2, a benchmark designed to probe skills associated with human-like generalization. Here, the nuance matters: Anthropic notes that the 60.4% figure is achieved at high effort, and that scores shown in its chart correspond to max effort settings—details that can materially affect comparisons.
Even so, the ARC-AGI-2 result appears strong against many peers in the same “general-purpose” bracket, while still trailing top-end entries like Opus 4.6 and other frontier results reported on public leaderboards (including Gemini 3 Deep Think and a refined GPT-5.2 variant).
What it means for users
For most people, Sonnet 4.6’s impact won’t be abstract benchmark math—it’ll be whether the model is more dependable in the boring, daily moments: following instructions without drift, staying consistent across long tasks, and making fewer avoidable mistakes in code and tool-driven workflows.
If the 1M-token beta context window holds up under real workloads, it could also shift how teams think about “AI-assisted” development—less about chatting over snippets, more about working with whole repositories and projects in a single, continuous thread.