Claude Sonnet 5 is Anthropic's mid-tier model, released June 30, 2026 as the successor to Sonnet 4.6. Anthropic frames it as the most agentic Sonnet the company has shipped: a model meant to plan a task, use tools like a browser or a terminal, and carry a multi-step job forward with less step-by-step supervision than earlier Sonnet generations needed. Anthropic's own positioning is that Sonnet 5 closes much of the gap to the flagship Opus line while staying in the Sonnet tier.
Where it shines
The model is built around sustained tool use rather than single-turn chat. It supports function calling, structured JSON output against a schema, PDF input, and vision, and it exposes a reasoning mode so it can work through a problem before answering. That combination fits agentic workflows well: a task that requires reading a document, calling a tool, checking the result, and deciding what to do next is exactly the shape Anthropic says this release targets. Anthropic reports 78.5% on OSWorld-Verified, its computer-use benchmark, and 46.8% on Humanity's Last Exam when the model is allowed to use tools — both are Anthropic's own published figures, not independently reproduced numbers, so they describe the vendor's account of the model's progress rather than a neutral third-party result.
Under the hood
Sonnet 5 carries a 1,000,000-token context window with a maximum synchronous output of 128,000 tokens per response. Output is text-only — this generation of Sonnet does not generate images or audio. Prompt caching is supported, which matters for agentic loops that resend a large, mostly-unchanged context (a codebase, a long document, a tool schema) across many turns of the same session.
Where it falls short
Anthropic's own announcement of this release does not state a knowledge cutoff date, so anyone relying on the model for very recent events or newly released libraries should verify rather than assume currency. The model's strengths are concentrated in tool-using, multi-step work; for short single-turn questions or simple classification, the agentic machinery is overhead rather than benefit, and a lighter model in the same lineup will usually answer just as well. Anthropic's own benchmark claims, since they come from the vendor testing its own model, are also worth treating as directional rather than settled — the scale of an improvement over Sonnet 4.6 is Anthropic's characterization until it is checked against independent workloads.
When to pick it
Sonnet 5 fits coding agents, research agents, and any workflow where the model needs to keep working across several tool calls without a human re-prompting it at every step — debugging a failing test suite, working through a multi-file refactor, or running a browse-and-summarize task end to end. It also fits workloads that need a genuinely large context window alongside tool use, such as reviewing a large codebase or a long document set in one session.
For work that does not need that autonomy — short-form generation, simple extraction, single-call classification — the agentic overhead is not doing anything useful, and a smaller, faster model will typically be the more sensible default.
Alternatives worth comparing
Within Anthropic's own lineup, Claude Opus 5 sits above Sonnet 5 as the flagship for the most complex agentic and enterprise work, and Claude Fable 5 sits above that as Anthropic's most capable widely released model — both are options if a workload turns out to need more headroom than Sonnet 5 provides. Anthropic's own positioning frames Sonnet 5 as approaching Opus-class performance on many tasks, which makes it worth starting there before reaching for a larger model, and only stepping up if the workload specifically demonstrates it needs the extra capability.
Deployment notes
Because Sonnet 5 is built around long, tool-heavy sessions, integrations should expect and handle multi-turn tool-call loops rather than a single request/response exchange — timeouts, retry logic, and tool-result formatting all need to account for a task that may take several rounds to complete. The 128,000-token output ceiling is generous but finite; a workflow that expects one call to return an entire large generated artifact should confirm it fits under that cap before relying on it in production.