Matthew Berman Reviews Claude Sonnet 4.6: Benchmarks, Computer Use and Safety

Anthropic has introduced Claude Sonnet 4.6 as a workhorse model, with reported improvements in coding, tool use, agentic work and computer use. At launch in February 2026, it offered a one-million-token context window in beta and became the default on Claude’s Free and Pro plans. I am excited about a fast model aimed at practical knowledge work.

The announced API price starts at $3 per million input tokens and $15 per million output tokens, the same starting rates as Sonnet 4.5. Anthropic emphasizes consistency and instruction following, as well as tasks such as building presentations and working with spreadsheets through products including Claude Cowork and Claude Code.

Anthropic’s OSWorld chart shows Sonnet 4.5 at 61.4% and Sonnet 4.6 at 72.5% on OSWorld-Verified. The benchmark asks an agent to complete practical tasks in a computer environment. Earlier points on the same chart use the original OSWorld rather than the later verified version, so the entire time series is not measured identically.

Computer use works by having a model inspect a screen and act through a virtual mouse and keyboard, rather than relying on a custom connector for every task. That makes it broadly useful, but also exposes it to prompt injection: a malicious instruction hidden in material the model reads may try to redirect its behavior.

Anthropic says its safety evaluations show substantially better resistance to prompt injection than Sonnet 4.5 and performance similar to Opus 4.6. These are evaluation results, not a guarantee that attacks cannot succeed. I think users should consider that risk when giving agents access to personal data or actions with real-world consequences.

Anthropic’s benchmark table reports terminal coding rising from 51.0% for Sonnet 4.5 to 59.1% for 4.6, and SWE-bench Verified from 77.2% to 79.6%. OSWorld-Verified rises from 61.4% to 72.5%. On its tool-use tests, retail improves while telecom remains close to its previous score.

The table also reports MCP-Atlas scaled tool use rising from 43.8% to 61.3%. On Humanity’s Last Exam, the score moves from 17.7% to 33.2% without tools and from 33.6% to 49.0% with tools. I would have expected an even larger tool-assisted jump, but these are the results shown.

For agentic financial analysis the reported score is 63.3%, and on the GDPval-AA office-task comparison Sonnet 4.6 has an Elo of 1,633, ahead of Opus 4.6’s 1,606 in the displayed table. Those are benchmark-specific results. I interpret them as evidence that Anthropic is targeting real-world knowledge work, though Opus leads many other rows.

The chart reports ARC-AGI-2 increasing from 13.6% to 58.3%, with Opus 4.6 at 68.8%. GPQA Diamond and visual reasoning also improve, while multilingual question answering is roughly unchanged. These different tests measure different capabilities and do not imply the same improvement across every use.

Vending-Bench Arena simulates running a vending business: managing inventory, purchases and reordering to improve the balance. The shown simulation ends around $5,500 for Sonnet 4.6 versus roughly $2,000 for 4.5 after more than 350 simulated days. Anthropic attributes the newer model’s result to early capacity investment followed by a late shift toward profitability; this chart is a simulation, not a real vending-machine profit record.

The launch also describes adaptive reasoning, API context compaction in beta, and updates to web search, code execution, memory and tool calling. Anthropic says the Claude in Excel add-in supports MCP connectors. These are launch-time product claims and availability can change.

The system card says Sonnet 4.6 was deployed under Anthropic’s AI Safety Level 3 standard. I outline Anthropic’s levels as a way to understand the safeguards and risk categories. Deployment under that standard describes the company’s risk-management framework; it does not establish that every dangerous capability is present or every risk is eliminated.

GDPval-AA, developed by Artificial Analysis using tasks from OpenAI’s GDPval dataset, covers 44 occupations across nine industries. The tasks mirror professional outputs such as documents, slides, diagrams and spreadsheets. Sonnet 4.6’s higher Elo than Opus 4.6 on this particular comparison supports my view that the model is aimed at knowledge work, but does not prove it is better on every professional task.

Anthropic’s system card concludes that Sonnet 4.6 does not cross its AI R&D-4 or CBRN-4 capability thresholds. The former concerns fully automating an entry-level remote Anthropic researcher’s work; the latter concerns high-consequence chemical, biological, radiological or nuclear misuse. These are Anthropic’s internal definitions and assessment.

The card also says confidently ruling out those thresholds is becoming harder, because evaluation results approach high capability levels and parts of the thresholds are difficult to measure. Anthropic says it has proactively applied mitigations associated with AI R&D-4. I find the uncertainty important: assessing capability can become harder even when a threshold has not been judged crossed.

My takeaway is that Sonnet 4.6 is positioned as a knowledge-work model, and its benchmark results narrow the apparent gap with Opus on some tasks. People online speculate about how Anthropic named the model or what it was originally intended to be, but I do not know the details. I plan to test it further.