Anthropic Releases Claude Sonnet 5, Cuts Opus 4.8 Price Gap on Knowledge Work
Anthropic released Claude Sonnet 5 on June 30, 2026, a midsize model with expanded agentic capabilities. As of today, it replaces Sonnet 4.6 as the default for both free and Pro plan users.
Benchmarks
On Anthropic's agentic coding benchmark, Sonnet 5 scores 63.2%. Sonnet 4.6 scored 58.1%. Opus 4.8 scored 69.2%.
The knowledge work benchmark reverses the ranking. Sonnet 5 slightly outperforms Opus 4.8 there. One benchmark up, one benchmark down, depending on what you care about.
Hallucination and sycophancy rates are lower than Sonnet 4.6. Refusal of malicious requests and resistance to prompt-injection attacks are improved. Dangerous cybersecurity task capability is lower than Opus models, which Anthropic lists as a safety property.
Pricing and the August 31 Cliff
Through August 31, 2026: $2 per million input tokens, $10 per million output tokens. Starting September 1: $3 and $15.
The two-month introductory window gives teams time to evaluate before committing to agentic workloads at the higher rate.
At current pricing, Sonnet 5 is cheaper than Opus 4.8, OpenAI's GPT-5.5, and Google's Gemini 3.1 Pro. Gemini 3.5 Flash, released in May 2026, still undercuts it.
What Sonnet 5 Actually Adds
Browser and terminal tool use are the headline agentic additions. These put it closer to what the competition has been shipping. OpenAI previewed GPT-5.6 Sol the week of June 23, describing it as their most agentic model with subagent task splitting. Google shipped Gemini 3.5 Flash in May with agentic planning and building capabilities.
Sonnet 5 is not the top of Anthropic's lineup on coding. It is, for now, the cheapest way to run agents against their models with meaningful safety constraints baked in.
The September Math
The practical question is whether Sonnet 5's knowledge work performance closes enough of the Opus 4.8 gap to justify the switch on long-running agentic jobs. At the introductory rate, probably yes for most use cases. At $3/$15, the calculation depends on how much of your workload lands in knowledge work versus coding.
Source: Techcrunch