Overview
Claude Sonnet 5 is Anthropic's mid-tier flagship, released on June 30, 2026, and pitched as "the best combination of speed and intelligence." It is Anthropic's most agentic Sonnet yet: it plans multi-step work, drives browsers and terminals, and carries out autonomous tasks at a level that needed a larger, pricier model only months earlier. It substantially narrows the gap to Opus 4.8 on agentic coding, tool-augmented reasoning, and computer use, while staying priced well below it.
Sonnet 5 is now the default model behind Anthropic's Free and Pro plans, and is also available on Max, Team, and Enterprise plans, in Claude Code, and through the Claude API (claude-sonnet-5).
Key capabilities
| Dimension | Detail |
|---|---|
| Context window | 1,000,000 tokens (the only tier — no smaller variant, no long-context surcharge) |
| Max output | 128,000 tokens (up to 300K on the Batch API with the output-300k beta) |
| Input modalities | Text, image |
| Output modalities | Text |
| Tools | function calling, structured outputs, streaming, prompt caching, web search, code execution, bash, text editor, computer use, MCP |
| Reasoning | Adaptive Thinking only, on by default; effort levels, defaulting to high on the Claude API and Claude Code |
Adaptive thinking now runs on every request unless you explicitly disable it; manual thinking budgets (
budget_tokens) are no longer available on this model — use theeffortparameter instead. Sonnet 5 also switches to the newer tokenizer shared with Opus 4.7 and later: the same text produces roughly 30% more tokens than on Sonnet 4.6, so recount prompts and revisitmax_tokensbudgets when migrating. Prompt caching keeps the same multipliers as Sonnet 4.6 — cache reads at 0.1x, 5-minute writes at 1.25x, and 1-hour writes at 2x the base input rate. See live pricing in the model catalog.
Benchmarks
Sonnet 5's benchmark story is closing the distance to Opus 4.8 on agentic coding, tool-augmented reasoning, and computer use, while comfortably outscoring Sonnet 4.6 across every axis Anthropic reports: SWE-Bench Pro (63.2%), Terminal-Bench 2.1 (80.4%), Humanity's Last Exam with tools (57.4%), OSWorld-Verified (81.2%), and GDPval-AA v2 (1618).
The chart above is Anthropic's own launch comparison, lining Sonnet 5 up against Sonnet 4.6 and Opus 4.8 (shown for reference) across all five evaluations.
Agentic coding
Terminal-Bench 2.1 nearly matches Opus 4.8
SWE-Bench Pro and Terminal-Bench 2.1 both grade a model inside real repositories and shells, not on isolated code generation.
SWE-Bench Pro and Terminal-Bench 2.1 both grade a model inside real repositories and shells rather than on isolated code generation. Sonnet 5 reaches 63.2% on SWE-Bench Pro, a 5.1-point jump over Sonnet 4.6, landing about 6 points behind Opus 4.8 at a fraction of the cost. On Terminal-Bench 2.1 the gap to Opus 4.8 nearly disappears: Sonnet 5's 80.4% sits within 2.3 points of Opus 4.8's 82.7%, and 13.4 points above Sonnet 4.6.
Agentic search: cost vs. pass rate
On BrowseComp, an agentic web-search benchmark, Anthropic plots pass rate against actual dollar cost per task across five effort settings (low/med/high/xhigh/max). Sonnet 5's curve tracks almost on top of Opus 4.8's across the whole range — within a couple of points at every effort tier, for a noticeably lower cost per task. Sonnet 4.6's curve sits well below both: even at its priciest (max) setting, it falls short of what Sonnet 5 reaches at high effort, at several times the price.
Computer use: cost vs. pass rate
The same pattern holds on OSWorld-Verified. Sonnet 5 tracks just behind Opus 4.8 at every effort level while costing less per task — its max-effort score edges past Opus 4.8's own low-effort score at a fraction of the price. Sonnet 4.6's curve sits clearly below both, needing its priciest setting just to approach what Sonnet 5 gets at its cheapest.
Safety and alignment
On Anthropic's automated behavioral audit (1–10 scale, lower is better), Sonnet 5 scores 2.53, down from Sonnet 4.6's 2.89, though still a step above Opus 4.8 (2.10) and the specialized Mythos Preview (1.95). Sonnet 5 is also the first Sonnet-tier model to ship real-time cybersecurity safeguards: requests touching prohibited or high-risk cybersecurity topics can be refused, and those refusals come back as a normal HTTP 200 response with stop_reason: "refusal" rather than an error — worth handling explicitly if your integration branches on error codes.
When to use it
- Cost-sensitive agentic coding and terminal automation: near-Opus SWE-Bench Pro and Terminal-Bench scores at a fraction of the cost per task.
- Long-running autonomous agents: browser, terminal, and MCP workflows where the cost curve matters as much as the ceiling.
- General-purpose default: it's the default model behind Free and Pro plans, and a straightforward upgrade for anything already running on Sonnet 4.6.
- When to pick a sibling instead: reach for Opus 4.8 when a task needs the highest ceiling regardless of cost, or Haiku 4.5 for latency-critical, simpler tasks.
CrossModel exposes Claude Sonnet 5 through both Anthropic-compatible /v1/messages and OpenAI-compatible /v1/chat/completions. Current pricing is available in the model catalog.