Sonnet 4.6 just made Opus look expensive. That changes everything.

Anthropic today launched Claude Sonnet 4.6, a mid-tier model that internal testing shows developers prefer to the company's most capable model from November 2025, Opus 4.5, 59 percent of the time. The figure, drawn from early testing in Claude Code, lands at a moment when the frontier AI market is dividing between expensive flagship models and cheaper alternatives that often lag in reasoning depth. GPT-5.6 just made every dollar in AI count harder

Sonnet 4.6 aims to collapse that distinction. Priced at $3 per million input tokens and $15 per million output tokens, identical to Sonnet 4.5, the model offers a 1 million token context window in beta, improved computer use skills, and across-the-board benchmark gains that Anthropic says approach Opus-level intelligence. The company's safety evaluations rated it as safe or safer than recent Claude releases, noting a broadly warm, honest, prosocial, and at times funny character and no high-stakes alignment concerns.

Benchmarking the gap

Anthropic released detailed benchmark comparisons against GPT-5.2, Gemini 3 Pro, and its own Opus 4.6 and Sonnet 4.5 models. On OSWorld, the standard benchmark for AI computer use, Sonnet 4.6 scored 47.5 percent, a jump from Sonnet 4.5's 27.3 percent and slightly ahead of Opus 4.6's 45.6 percent. On Terminal-Bench 2.0, which tests agentic coding tasks, Sonnet 4.6 reached 89.6 percent, topping GPT-5.2's 88.1 percent and Opus 4.6's 87.9 percent. On SWE-bench Verified, a software engineering benchmark, it scored 79.1 percent, outperforming GPT-5.2's 73.0 percent and trailing Opus 4.6's 81.5 percent by a narrow margin. Your AI assistant forgets you every morning. This…

On OfficeQA, which measures enterprise document comprehension, Sonnet 4.6 matched Opus 4.6 exactly at 90.1 percent. The model also posted strong results on BrowseComp (67.3 percent), ARC-AGI-2 (66.8 percent with max effort), and MMMU-Pro (83.0 percent). On Humanity's Last Exam, a grueling benchmark released in January 2025, Sonnet 4.6 with thinking and tools scored 54.1 percent, behind Opus 4.6's 79.7 percent but ahead of Sonnet 4.5's 46.9 percent.

The numbers paint a clear picture: Sonnet 4.6 closes the gap to Opus on most practical tasks while leaving a significant margin only on the hardest reasoning evaluations. For the majority of coding, document analysis, and agentic workflows, the difference is negligible.

Computer use matures

Anthropic first introduced general-purpose computer-using models in October 2024, then described as still experimental, at times cumbersome and error-prone. Sonnet 4.6 marks the third generation of that capability, and the OSWorld scores reflect sixteen months of steady improvement. The model navigates real software environments, Chrome, LibreOffice, VS Code, by clicking a virtual mouse and typing on a virtual keyboard, with no special APIs or connectors.

Early customer testimonials published by Anthropic describe human-level performance on tasks like navigating complex spreadsheets, filling out multi-step web forms, and coordinating across multiple browser tabs. Box reported a 15 percentage point improvement over Sonnet 4.5 on heavy reasoning Q&A. An insurance benchmark hit 94 percent, the highest the customer had tested for computer use. Rakuten AI called Sonnet 4.6's iOS code the best it had tested, noting better spec compliance, architecture, and modern tooling usage.

The company acknowledged that computer use still lags behind the most skilled humans and carries risks from prompt injection attacks, where malicious actors hide instructions on websites to hijack the model. Anthropic said Sonnet 4.6 shows major improvement in resistance to such attacks, performing similarly to Opus 4.6 on internal safety evaluations. The subtle trap waiting for AI agents in production

The pricing paradox

The most striking implication of Sonnet 4.6 is not technical but economic. Opus 4.6 costs $15 per million input tokens and $75 per million output tokens, five times the output cost of Sonnet 4.6. For a developer running thousands of agentic coding sessions per month, the cost difference is substantial.

Anthropic's own data suggests that users prefer Sonnet 4.6 to Opus 4.5 across a wide range of tasks, including codebase refactoring, multi-step planning, and bug detection. The company explicitly notes that Sonnet 4.6 provides a viable alternative if you are a heavy Opus user. The recommendation comes from the company that sells Opus.

For teams running agentic coding at scale, the metric that matters is not raw benchmark scores but cost per resolved issue. Sonnet 4.6 matches Opus on bug detection rates at one-fifth the output cost, meaning teams can run more reviewers in parallel without increasing their budget. The economics of scale now favor the mid-tier model for the first time.

Product updates and availability

Sonnet 4.6 is available immediately on all Claude plans, Free, Pro, Max, Team, and Enterprise, and on Claude Code, the API, and all major cloud platforms. The free tier now includes file creation, connectors, skills, and context compaction by default. The model supports both adaptive thinking and extended thinking, as well as context compaction in beta, which automatically summarizes older conversation context to extend effective session length. Anthropic launches Claude Tag on Slack for team-wide AI…

On the API side, Anthropic made web search and fetch tools generally available, with automatic code execution to filter and process results. Memory, programmatic tool calling, tool search, and tool use examples also moved to general availability. For Excel users, the Claude add-in now supports MCP connectors, allowing the assistant to pull data from S&P Global, LSEG, Daloopa, PitchBook, Moody's, and FactSet without leaving the spreadsheet.

The developer endpoint uses claude-sonnet-4-6 as the model identifier. Anthropic recommends migrating from Sonnet 4.5 and exploring across the thinking effort spectrum to find the optimal balance between speed and reliability for each use case.

What it means for the market

Sonnet 4.6 represents a structural shift in how AI labs are positioning their model stacks. For the past two years, the standard advice has been: use the cheapest model that works, and reach for Opus or GPT-5 only when absolutely necessary. That calculus worked because the gap between mid-tier and flagship was large enough to justify the premium. The specialization revolution: how smaller models are…

That gap is now shrinking. Anthropic's own internal user preference data shows developers voluntarily choosing a model that costs one-fifth as much, on tasks where they previously would have paid for the top tier. If that pattern holds at scale, it may force a broader reconsideration of how the industry prices frontier intelligence. If a mid-tier model can do what the flagship did six months ago, the flagship must differentiate on tasks that few real-world users encounter.

Anthropic is not unaware of this. The company notes that Opus 4.6 remains the strongest option for tasks that demand the deepest reasoning, such as codebase refactoring, coordinating multiple agents in a workflow, and problems where getting it just right is paramount. But the list is shorter than it was a year ago, and it grows shorter with each Sonnet release. Anthropic Launches Claude Sonnet 5: A New Frontier in…