Claude Sonnet 4.6 Official Release: Coding, Computer Use, and Agent Benchmarks

One-sentence takeaway

Anthropic's release data positions Sonnet 4.6 as a lower-cost Opus-level candidate: it performs strongly on SWE-bench, OSWorld, OfficeQA, financial agents, and long contexts, but Opus 4.6 is still worth considering for extremely deep reasoning, complex search, and difficult refactoring.

Test environment

• Model: Claude Sonnet 4.6, API entry point claude-sonnet-4-6; the 1M-token context window is in beta.

• Pricing: The release page states a starting price of $3/$15 per million input/output tokens, the same as Sonnet 4.5.

• Benchmarks/workflows: SWE-bench Verified, OSWorld-Verified, Vending-Bench Arena, OfficeQA, Finance Agent v1.1, GDPval-AA, and Claude Code user preferences; the harness differs across metrics.

• Version boundary: Anthropic specifically notes that Sonnet 4.5 and earlier used the original OSWorld, while Sonnet 4.5 onward uses OSWorld-Verified, so results cannot be compared directly across versions.

Inputs/configuration

The release page does not disclose the complete prompts, sampling parameters, number of repetitions, or per-question trajectories for all benchmarks. Claude Code user preferences come from early testing; Vending-Bench Arena is a simulated long-horizon management task; OSWorld-Verified uses simulated computers, mouse and keyboard, and real software scenarios.

Results data

• In Claude Code, users preferred Sonnet 4.6 over Sonnet 4.5 about 70% of the time; versus Opus 4.5, users preferred it about 59% of the time.

• In the data cited on the release page, Sonnet 4.6 matched Opus 4.6 on OfficeQA; Anthropic says it first reinvested in capability in Vending-Bench Arena, then turned toward profitability, and finished ahead.

• Anthropic's customer/partner feedback says that Box improved by 15 percentage points over Sonnet 4.5 on reasoning-intensive question answering; one insurance benchmark reached 94%, but these cases do not disclose a complete, standardized methodology.

• The release page says Sonnet 4.6 is significantly better at computer use than the previous generation and approaches a human-level experience on multi-tab spreadsheets and multi-step forms; it also acknowledges that the model still trails skilled humans.

Conclusion

Sonnet 4.6 is suitable for routing large volumes of routine coding, browser/desktop automation, enterprise document, and financial analysis tasks away from Opus to a cheaper default model; for deep refactoring, multi-agent coordination, and complex agentic search, run a Sonnet/Opus routing evaluation.

Limitations

• The results are self-reported by the vendor, and many depend on Anthropic's own scaffolds, early user preferences, or customer cases; complete raw data and variance are not disclosed.

• Commonly cited figures such as 79.6% on SWE and 72.5% on OSWorld should be checked against the corresponding configurations in the system card or release page; do not treat all benchmark results as a single leaderboard.

• The 1M context window is in beta, and long-context pricing and platform limits may differ.

• Computer use faces risks such as prompt injection; the release page only describes improvements in safety evaluations, so production deployments still require isolation, permissions, and handling of web content as untrusted.

Reproduction steps

1. Fix the claude-sonnet-4-6 snapshot, effort, thinking, tool permissions, and context-window configuration.

2. Build a consistent task set covering real code issues, browser forms, enterprise PDFs/spreadsheets, and financial analysis, and save all tool trajectories.

3. Run Sonnet 4.6 and Opus 4.6 separately, reporting accuracy, resolved status, tool calls, latency, tokens, cost, and prompt-injection failures.

4. Treat the official figures as baselines rather than reproduction results, and explain the differences between your harness and Anthropic's harness.