Anthropic releases Claude Sonnet 5 at $2/1M input tokens, 63.2% agentic coding benchmark

Anthropic released Claude Sonnet 5 on Tuesday, positioning it as a cost-effective option for running autonomous agents. The model is priced at $2 per million input tokens and $10 per million output tokens through August 31, after which input pricing rises to $3 per million tokens.

Performance metrics

On agentic coding benchmarks, Sonnet 5 scores 63.2%, compared to Opus 4.8's 69.2% and its predecessor Sonnet 4.6's 58.1%. On knowledge work tasks, Sonnet 5 slightly outperforms Opus 4.8, according to Anthropic. The company claims the model can "make plans, use tools like browsers and terminals, and run autonomously" at levels previously requiring larger models.

Daniel Shepard, senior engineer at Zapier, reported that Sonnet 5 completed a two-part Salesforce automation task end-to-end, whereas previous versions would stall midway. The model also reportedly checks its own output without explicit prompting.

Pricing comparison

Claude Sonnet 5 undercuts several competitors:

• Cheaper than Opus 4.8 (pricing not disclosed in source)
• Cheaper than OpenAI's GPT-5.5 (pricing not disclosed)
• Cheaper than Gemini 3.1 Pro (pricing not disclosed)
• More expensive than Gemini 3.5 Flash (pricing not disclosed)

The model becomes the default for Claude's free and Pro plans starting Tuesday.

Safety improvements

Sonnet 5 shows lower rates of "undesirable behaviors" including cooperation with misuse, deception, hallucination, and sycophantic responses compared to Sonnet 4.6. It demonstrates improved performance at refusing malicious requests and resisting prompt injection attacks.

However, Anthropic notes it does not match Opus 4.8 or Claude Mythos Preview for handling misaligned behavior. The company states it "has a much lower ability to perform dangerous cybersecurity tasks than our current Opus models."

Fabian Hedin, co-founder of Lovable, stated the model "refuses unsafe requests cleanly and consistently," emphasizing the importance of models that "know when to say no."

Market context

The release follows similar agentic-focused launches from competitors. OpenAI released GPT-5.6 Sol last week with subagent capabilities for autonomous tasks. Google launched Gemini 3.5 Flash in May, also emphasizing agentic capabilities with minimal human oversight.

What this means

Agentic capability is now table stakes across model tiers, shifting competition to price and reliability. Anthropic is explicitly positioning Sonnet 5 as the cost-efficient option between its budget and premium offerings, betting that developers will trade small performance gaps for significant cost savings. The 5.2 percentage point gap between Sonnet 5 and Opus 4.8 on coding tasks may prove negligible for many production use cases, making the pricing the determining factor. The emphasis on safety features suggests Anthropic is responding to enterprise concerns about autonomous agents operating without human oversight.