Bold claim: Claude Sonnet 4.6 is Anthropic’s most capable Sonnet yet, and it arrives with a generous 1 million token context—plus claims of stronger safety with fewer hallucinations and less sycophancy. But here's where it gets controversial: does a longer context and safety brag translate into real-world reliability across tasks, or are benchmarks painting a more optimistic picture than everyday performance?
Overview
Anthropic has released Claude Sonnet 4.6, the latest in its Sonnet line, following the Feb. 5 introduction of Claude Opus 4.6. The company positions Sonnet 4.6 as its strongest Sonnet to date, boasting a beta context window of 1 million tokens. In internal safety assessments, Anthropic reports a lower tendency to hallucinate and less tendency toward requiring flattery or alignment with user demands.
Key capabilities
- Coders gain noticeably improved results: Anthropic highlights that Sonnet 4.6 strengthens Claude’s coding abilities for a broader audience of developers.
- Benchmark performance places Sonnet 4.6 at the top among Anthropic models for agentic financial analysis and office tasks, outperforming competitors such as Google’s Gemini 3 Pro and OpenAI’s GPT-5.2 in those areas. It also beats Anthropic’s own Opus 4.6 on several tasks.
- In the company’s system card, Sonnet 4.6 shows improvements on benchmarks like Humanity’s Last Exam, though Opus 4.6 still leads in some metrics.
How to access Claude Sonnet 4.6
- Availability: Sonnet 4.6 is the default model on Claude.ai and Claude Cowork for both free and Pro users. It is also accessible via Anthropic’s API and through major cloud platforms.
- Usage limits: Free users face usage limits tied to demand, with resets every five hours. Higher limits are available at the same price points as the prior model.
- Pricing: Claude Pro is $20 per month (or $17 per month with annual payment). API pricing starts at $3 per million input tokens and $15 per million output tokens.
What the benchmarks mean
- Benchmark highlights include GPQA Diamond (89.9%), ARC-AGI-2 (58.3%), MMMLU (89.3%), SWE-bench Verified (79.6%), and Humanity’s Last Exam with tools (49.0%) / without tools (33.2%).
- Industry note: Pace, an AI-powered insurance company, reported that Sonnet 4.6 scored the best among Claude models on its complex insurance-use benchmark.
Cost versus capability
- Sonnet 4.6 offers higher perceived capability at a lower relative cost than some Opus models (e.g., 3 dollars per million input tokens and 15 dollars per million output tokens, compared with Opus 4.6 at 5/25).
Bottom line
Claude Sonnet 4.6 marks a notable step forward in both capability and accessibility within Anthropic’s lineup, especially for developers who rely on coding and complex reasoning tasks. However, as with any model release anchored by benchmark performance, it’s wise to test in your own workflows to verify real-world gains and to watch for variability across use cases. Do you value higher context length and stronger safety alignment more for your projects, or do you prioritize other factors like latency, cost, and domain-specific performance? Share your priorities and experiences in the comments.
