Anthropic promises ‘Opus-level’ reasoning with new Claude Sonnet 4.6 model – and all at a far lower cost

Anthropic has announced Claude Sonnet 4.6, its latest ‘all rounder’ AI model which it says can achieve results previously only possible with its most sophisticated Opus family of models – and all at a fraction of the price.

Claude Sonnet 4.6 is intended as an extremely capable model for coding and computer use, with tangible performance improvements across real-world, economically-valuable tasks.

Anthropic stressed that these new capabilities are paired with API costs low enough for businesses to reasonably use it as a backend model for high-performance AI agents, or to upgrade the outputs of their existing Claude Sonnet deployments at no extra cost.

Indeed, via the Claude API, the new model is accessible at the same price as Claude Sonnet 4.5: $3 per million input tokens and $15 per million output tokens.

This compares to the Claude Opus 4.6 pricing of $5 per million input tokens and $25 per million output tokens.

But industry experts describe its performance as significantly improved compared to the previous Sonnet release.

"The performance-to-cost ratio of Claude Sonnet 4.6 is extraordinary—it’s hard to overstate how fast Claude models have been evolving in recent months,” said Michele Catasta, president at Replit.

“Sonnet 4.6 outperforms on our orchestration evals, handles our most complex agentic workloads, and keeps improving the higher you push the effort settings."

Claude Sonnet 4.6 benchmarks: How does it compare?

In the popular coding benchmark SWE-bench Verified, Claude Sonnet 4.6 scored 79.6%, comparable to the score of other frontier models.

Perhaps more importantly, real-world testing found developers in Claude Code preferred Claude Sonnet 4.6 outputs around 70% of the time – and even found its code outputs more thorough than Claude Opus 4.5, which until recently had been Anthropic’s most powerful model.

The new model also boasts a one million token context window, in beta, enough to hold entire code bases or multiple complex documents for context.

Claude Sonnet 4.6 also demonstrates a significant leap in agentic tool use and computer use compared to its predecessor.

For example in OSWorld-Verified, a benchmark intended to test how well AI agents can use web and desktop apps, Claude Sonnet 4.6 scored 72.5%, a marked improvement on the 61.4% scored by the previous Sonnet model and even more than the 66.3% achieved by Claude Opus 4.5.

Crucially, Claude Sonnet 4.6’s score in this benchmark is only 0.2% behind that of Anthropic’s flagship model Claude Opus 4.6, launched two weeks ago.

In the real-world office task benchmark GDBval-AA-Elo, Claude Sonnet 4.6 even outperforms Claude Opus 4.6, a feat it also achieves in the agentic financial analysis benchmark Finance Agent v1.1.

"Claude Sonnet 4.6 matches Opus 4.6 performance on OfficeQA, which measures how well a model can read enterprise documents (charts, PDFs, tables), pull the right facts, and reason from those facts,” said Hanling Tang, CTO of Neural Networks at Databricks.

“It’s a meaningful upgrade for document comprehension workloads."

How to access Claude Sonnet 4.6

Claude Sonnet 4.6 is accessible now via all channels to access Anthropic’s models, including the Claude API, Claude Cowork, Claude subscription plans, and via public cloud platforms.

In its announcement, Anthropic confirmed that Claude Sonnet 4.6 is now the default model on the Claude web app, as well as Claude Cowork, and has upgraded its free-tier capabilities to include file creation, skills, and code compaction to optimize long conversations.

The new model launches alongside new support for MCP connectors inside Claude in Excel, the company revealed.

This allows users to pull external context into Claude from within their spreadsheet and to connect directly to tools such as Daloopa, FactSet, LSEG, Moody’s, Pitchbook, and S&P Global.