Claude Sonnet 4.6 is now Anthropic’s default Sonnet model across claude.ai and Claude Cowork, positioning the “mid-tier” Claude line closer to what previously required stepping up to Opus. The update lands with a broad set of capability claims—coding, computer use, long-context reasoning, agent planning, knowledge work, and design—and introduces a 1M token context window (beta).
Anthropic says Sonnet 4.6 keeps the same API pricing as Sonnet 4.5, starting at $3/$15 per million tokens, and is available on Free and Pro plans as the default model. It’s also available via Claude plans, Claude Cowork, Claude Code, the API, and “all major cloud platforms.” Developers can select it in the API using claude-sonnet-4-6.
Coding: preference testing tilts toward Sonnet 4.6
In Claude Code preference testing, Anthropic reports users chose Sonnet 4.6 over Sonnet 4.5 about 70% of the time. Feedback highlighted a few recurring themes: better context reading before making edits, a tendency to consolidate shared logic instead of duplicating it, and fewer frustrations across longer sessions.
Interestingly, the same testing reported users preferred Sonnet 4.6 over Opus 4.5 (Anthropic’s November 2025 “frontier model”) 59% of the time, citing less overengineering, less “laziness,” improved instruction following, and fewer false claims of success or hallucinations.
Computer use: OSWorld gains and stronger prompt-injection resistance
Anthropic continues to push “computer use” as an automation path for legacy and specialized software that doesn’t expose modern APIs. Sonnet 4.6 builds on Anthropic’s earlier computer-using model work introduced in October 2024, with progress tracked on OSWorld—a benchmark that runs tasks across real apps like Chrome, LibreOffice, and VS Code in a simulated environment.
Alongside capability improvements, Anthropic flags prompt-injection as a key risk for computer-using models. Its safety evaluations (linked below) indicate Sonnet 4.6 is a major improvement over Sonnet 4.5 in prompt-injection resistance, performing similarly to Opus 4.6. Anthropic also points to guidance in its docs on mitigating jailbreaks and strengthening guardrails:
Mitigate jailbreaks
Long context and planning: 1M tokens (beta), plus compaction
The headline context upgrade is the 1M token context window in beta, framed as large enough for entire codebases, long contracts, or large sets of papers—and, crucially, as something Sonnet 4.6 can reason over effectively rather than merely store.
Anthropic also highlights long-horizon planning behavior it observed on Vending-Bench Arena, an evaluation that tests running a simulated business over time. Sonnet 4.6 reportedly adopted a strategy of investing heavily in capacity early, then pivoting to profitability later, outperforming Sonnet 4.5 under that setup.
On the developer platform, Sonnet 4.6 supports adaptive thinking and extended thinking, and includes context compaction (beta) to summarize older context as conversations approach limits.
Tooling updates on the API and beyond
Anthropic also bundles several platform notes with the model release:
- Claude’s web search and fetch tools now automatically write and execute code to filter and process search results, referencing: Improved web search with dynamic filtering
- Code execution, memory, programmatic tool calling, tool search, and tool use examples are now generally available, with implementation notes here: Providing tool use examples
For Claude in Excel, Anthropic says the add-in now supports MCP connectors and can reuse existing Claude.ai connector setups automatically. This is available on Pro, Max, Team, and Enterprise plans: Claude in Excel
Safety evaluations and system card
As with other Claude releases, Anthropic links to its safety work and a system card covering capability and behavior details:
Claude Sonnet 4.6 system card