Claude Sonnet 5

Claude Sonnet 5 (API model ID claude-sonnet-5) is the Sonnet-tier model in Anthropic's Claude 5 generation, released 2026-06-30 as "the most agentic Sonnet model yet" — an upgrade to Sonnet 4.6 in reasoning, tool use, coding, and knowledge work (Introducing Claude Sonnet 5). Anthropic frames its performance as "close to that of Claude Opus 4.8, but at lower prices" (Introducing Claude Sonnet 5), and it sits below Claude Fable 5 and the invitation-only Claude Mythos line in the "Latest models comparison" table on Anthropic's own model docs (Models overview). It is the default model for Claude Free and Pro plan users at launch, also available on Max, Team, and Enterprise plans, in Claude Code, and via the API (Introducing Claude Sonnet 5).

Key specs

Values re-verified 2026-07-12 directly against Anthropic's live platform docs and the models.dev registry. The registry snapshot lists Sonnet 5's release date as 2026-06-29, one day earlier than Anthropic's own announcement and the AWS Bedrock model card (both say 2026-06-30); this page follows the vendor date and flags the one-day gap rather than silently picking one.

Spec	Value	Source
Context window	1,000,000 tokens	Models overview
Max output	128,000 tokens (Messages API); up to 300,000 on the Batch API with the output-300k-2026-03-24 beta header	Models overview
Pricing, introductory (through 2026-08-31)	$2.00 / MTok input, $10.00 / MTok output	Introducing Claude Sonnet 5; Models overview
Pricing, standard (from 2026-09-01)	$3.00 / MTok input, $15.00 / MTok output	Models overview
Prompt caching	Minimum cacheable prompt 1,024 tokens (same floor as Claude Opus 4.8 and Sonnet 4.6); cache read $0.20 / MTok (0.1x base input); 5-minute cache write $2.50 / MTok (1.25x); 1-hour cache write $4.00 / MTok (2x)	Prompt caching docs
Thinking	Adaptive thinking on by default — runs even when thinking is omitted; pass thinking: {type: "disabled"} to turn it off; manual {type: "enabled", budget_tokens: N} is rejected with a 400 error	Adaptive thinking docs
Effort levels	low / medium / high / xhigh / max; xhigh is new for the Sonnet tier with this release (previously Opus/Fable-only); effort defaults to high on the Claude API and Claude Code	Adaptive thinking docs; Models overview
Vision resolution	High-resolution tier: 2,576px long edge, up to 4,784 visual tokens per image — first Sonnet-tier model in this tier (Sonnet 4.6 stays on the standard 1,568px / 1,568-token tier)	Vision docs
Modalities	Input: text + image; output: text only	Bedrock model card
Knowledge cutoff	January 2026 (reliable and training-data cutoff both)	Models overview
Release date	2026-06-30	Introducing Claude Sonnet 5; Bedrock model card

Benchmark results

On the independent SWE-bench Verified leaderboard run by Vals AI's own harness, Claude Sonnet 5 resolves 79.6% of issues as of 2026-07-12 — behind Claude Fable 5 (95.0%) and Claude Opus 4.8 (88.6%) in the same Anthropic generation, and behind several non-Anthropic models on this specific leaderboard (Vals AI SWE-bench Verified leaderboard, as of 2026-07-12). This is the one benchmark figure on this page independently run by a third party against a public harness, rather than self-reported by Anthropic.

Benchmark	Sonnet 5	Comparator	According to
SWE-bench Pro	63.2%	—	Vellum.ai, citing Anthropic's System Card
Terminal-Bench 2.1	80.4%	—	Vellum.ai, citing Anthropic's System Card
OSWorld-Verified	81.2%	Sonnet 4.6 (updated eval methodology): 78.5%	Sonnet 5 figure: Vellum.ai; comparator: Anthropic announcement footnote (primary)
Humanity's Last Exam, with tools	57.4%	Sonnet 4.6 (updated grader): 46.8% with tools, 34.6% no tools	Sonnet 5 figure: Vellum.ai; comparator: Anthropic announcement footnote (primary)
GDPval-AA v2 (Elo-style knowledge-work score)	1,618	—	Vellum.ai, citing Anthropic's System Card
Firefox 147 exploit development (cyber eval)	0.0% complete exploits, 13.2% partial success	—	Vellum.ai, citing Anthropic's System Card

All Anthropic-reported rows as of the 2026-06-30 release. Note the OSWorld and HLE comparator numbers are Anthropic's own updated Sonnet 4.6 scores (the announcement's footnotes explain Anthropic changed the OSWorld-Verified evaluation method and the HLE grader model for this release), not Sonnet 4.6's original launch-day scores — a recurring methodology caveat for cross-generation vendor comparisons: confirm which run produced each number before treating a Sonnet-5-vs-Sonnet-4.6 delta as apples-to-apples.

One result outside Anthropic's own reporting: Cursor's internal CursorBench put Sonnet 5 at 57% task success, as cited by Vellum.ai — Cursor's own publication was not located directly in this research pass, so treat this as a single-lineage, unverified data point rather than a confirmed cross-lab result.

Agentic behavior notes

• Adaptive thinking runs by default, silently. Unlike Sonnet 4.6 (thinking off unless thinking: {type: "adaptive"} is set) and unlike Claude Fable 5 (thinking cannot be disabled at all), Sonnet 5 runs adaptive thinking automatically whenever the thinking parameter is omitted; code that relied on that omission meaning "no thinking" on Sonnet 4.6 gets adaptive thinking on Sonnet 5 unless updated to pass thinking: {type: "disabled"} explicitly (Adaptive thinking docs).
• xhigh effort is new for this tier. Anthropic's effort-level guidance describes xhigh as making Claude "always think deeply with extended exploration," and Sonnet 5 is the first Sonnet-tier model where it's available, alongside Claude Fable 5 and Claude Opus 4.8 (Adaptive thinking docs).
• Context editing is available, generically. Anthropic's context-editing docs (clearing old tool results and thinking blocks to manage a shrinking budget) state the feature is "available on all supported Claude models" without excluding Sonnet 5, but the docs' example code and per-model behavior tables (e.g. thinking-block clearing defaults) only enumerate models up through Sonnet 4.6 by name — Sonnet 5's specific defaults are not separately documented as of 2026-07-12 (Context editing docs). Practitioners relying on default tool-result-clearing behavior should verify it directly rather than assume Sonnet 4.6's table still applies.
• Not optimized for offensive cyber capability, per the benchmark table above: 0.0% complete exploits on the Firefox 147 vulnerability-exploitation eval, with only partial success at 13.2% (Vellum.ai, citing Anthropic's System Card) — Anthropic positions this as a deliberate design choice distinguishing Sonnet 5 from the invitation-only, cyber-focused Claude Mythos line rather than an incidental capability gap.
• Computer use on Bedrock uses tool version computer_20251124 (beta header computer-use-2025-11-24) (Bedrock model card).

Limitations

• Breaking API changes from Sonnet 4.6. temperature, top_p, and top_k return a 400 error on any non-default value on Sonnet 5 (this applies to every request, whether or not thinking is active); manual thinking: {type: "enabled", budget_tokens: N} is rejected outright rather than merely deprecated; and omitting thinking now runs adaptive thinking by default instead of thinking-off. Naive model-string swaps from Sonnet 4.6 can hard-fail or silently change latency and cost (Adaptive thinking docs); the sampling-parameter rejection is independently corroborated by Simon Willison.
• Tokenizer inflation can erase the headline price cut. See the Key specs warning above — Anthropic's own 1.0–1.35x figure and Willison's independent per-language measurements (up to 1.42x on English) both point the same direction: budget real per-task cost, not nominal per-token price, before deciding Sonnet 5 is cheaper than Sonnet 4.6 for a given workload.
• Vision spot-check miss. In one independent tester's informal image-recognition check, Sonnet 5 misidentified a pelican photo as a goose, generating "a white goose riding a bicycle" instead (Simon Willison). This is a single anecdotal data point (N=1), not a systematic vision-accuracy eval, and should not be read as a general capability finding.
• Most Sonnet-5-specific benchmark numbers on this page are vendor-reported and not independently re-verified against the primary System Card. See the note in Benchmark results — Anthropic's own comparison table is image-only and the source PDF could not be fetched in this research pass; the numbers above rest on one secondary outlet's citation of that PDF, not on this page's own primary read or a third-party leaderboard (the SWE-bench Verified row is the exception: that one is independently run by Vals AI).

Related

• Claude Opus 4.8 — the model Sonnet 5 is explicitly benchmarked against ("close to Opus 4.8 performance, at lower prices"); shares Sonnet 5's 1,024-token prompt-caching floor and 2,576px vision tier.
• Claude Fable 5 — the top tier of the same generation, positioned above both Sonnet 5 and Opus 4.8; unlike Sonnet 5, its adaptive thinking cannot be disabled at all.
• Claude Code — one of the primary surfaces where Sonnet 5 is deployed and set as a default.
• Context window, prompt caching, and adaptive thinking — the mechanics behind the Key specs table above.
• Computer use and context compaction — agentic-loop features relevant to long-horizon use of this model.
• SWE-bench Pro, Terminal-Bench, OSWorld, Humanity's Last Exam — the benchmark pages behind the Anthropic-reported scores above.