On June 30, 2026, Anthropic officially announced Claude Sonnet 5. The focus of this update is not on single-turn conversation scores, but on making Sonnet-level models better at planning, using browsers and terminals, and performing multi-step tasks continuously. Anthropic says its overall capabilities are close to Opus 4.8, but with lower operating costs.
Sonnet 5 is now available to all Claude plans and has become the default model for Free and Pro plans; Claude Code and Claude API are also available, with the API model named claude-sonnet-5.
The core change is "getting things done"
Sonnet 4.6 already supports writing code and calling tools, while Sonnet 5 shifts the focus to longer execution chains. The early test scenarios showcased by the official team include handling real code warehouses, debugging faults, calling multiple tools, and verifying results during the process.
This is more valuable for programming agents and business automation: what users truly care about is not whether the model can write a function, but whether it can track context, adhere to project agreements, complete tests, and continue moving forward when problems arise midway. However, "close to Opus 4.8" does not mean that all tasks are at the same level; the official statement is that high effort settings can match Opus on certain tasks.
Prices go down, but bills don't necessarily decrease proportionally
As of August 31, 2026, the API price for Sonnet 5's introductory period is $2 per million input tokens and $10 per million output tokens; Afterwards, prices recovered to $3 and $15. During the same period, Opus 4.8 was $5 and $25, so tasks requiring a large number of tool calls have clear cost margin.
But Sonnet 5 has switched to a new tokenizer. Anthropic reminds that the same input may result in about 1.0-1.35 times the original token, depending on the content type. When teams conduct migration evaluations, they should compare the total tokens, success rate, and number of human takeovers for a complete task, rather than just looking at unit price.
Who is most worth upgrading first?
If your workflow already relies on Claude Code, browser search, terminal commands, or long-link automation, Sonnet 5 is worth prioritizing for small-scale A/B testing. For ordinary chats, short text rewriting, and one-off Q&A, the value of new models may not be as obvious as in agent scenarios.
On the safety front, Anthropic reports that Sonnet 5 experiences fewer hallucinations and flattery than Sonnet 4.6, and its resistance to prompt injection has improved; But these are still the publishers' assessments. A safer approach before going live is to use the team's own codebase, permission boundaries, and failure samples for retesting, then decide whether to replace the existing model.
