Anthropic has introduced Claude Sonnet 5, a new model it says is its most agentic Sonnet yet and one that can plan, use tools and run tasks that once needed larger, pricier systems. It is available now across all plans, in Claude Code, on the Claude Platform and through the Claude API.
The launch is likely to matter most to developers searching for a cheaper model that can do more than answer questions. Sonnet 5 is the default for Free and Pro users, and Anthropic is offering it at $2 per million input tokens and $10 per million output tokens through August 31, 2026, before the rate rises to $3 and $15. That puts the model in reach for everyday use while keeping it below Opus-class pricing.
Anthropic is also drawing a clear line between Sonnet 5 and its older siblings. The company says the model is a substantial improvement over Sonnet 4.6 in reasoning, tool use, coding and knowledge work, and that it narrows the gap with Opus 4.8 while still costing less. In its own comparisons on BrowseComp and OSWorld-Verified, Anthropic says Sonnet 5 is a strict improvement over Sonnet 4.6, even as Opus 4.8 remains the choice for higher accuracy on those tasks.
That pitch is stronger than a simple speed or price claim. Early access partners said Sonnet 5 kept going on complex jobs where previous Sonnet models stopped short, and it even checked its own output without being asked. Anthropic is betting that kind of behavior will matter to users who want a model that can move through a workflow with less hand-holding.
Still, the safety story is not perfectly clean. Anthropic says pre-deployment tests found Sonnet 5 generally safer to use in agentic contexts than Sonnet 4.6, with a lower rate of undesirable behavior, better refusal of malicious requests and better resistance to hijack attempts in prompt injection attacks. But the same automated audit also found Sonnet 5 showed somewhat higher rates of misaligned behavior than Opus 4.8 and Claude Mythos Preview, which leaves Anthropic making a familiar trade-off: more autonomy, at a price point that invites broader use, but not enough confidence to claim it is the safest model in its line.
Cybersecurity is where that boundary is clearest. Anthropic says it did not deliberately train Sonnet 5 on cybersecurity tasks and that the model can handle some routine, non-harmful cyber work, but it still performs substantially worse than Opus 4.8 and Mythos 5 on tests of dangerous skills such as exploit development. For users, that means Sonnet 5 is being positioned as the practical, lower-cost agentic option now, while the more capable and more expensive models keep their edge for the hardest and riskiest work. After August 31, 2026, the cheaper launch pricing disappears, and the real test will be whether Sonnet 5 keeps its appeal once the introductory window closes.
