Claude Sonnet 5

Claude Sonnet 5

Quick answer: Claude Sonnet 5 is Anthropic's most capable Sonnet-class model, released 30 June 2026. It closes most of the gap to Opus 4.8 at Sonnet pricing ($3/$15 per million tokens), ships with adaptive thinking on by default (selectable effort from low to x-high), a 1M-token context window, and real-time cybersecurity safeguards. Anthropic reports 91.1% on GPQA Diamond and 39.6% on Humanity's Last Exam. API identifier: claude-sonnet-5.

At a Glance

Where Claude Sonnet 5 leads

• Near Opus 4.8 agentic performance at Sonnet-class price — $3/$15 per 1M tokens (introductory $2/$10 through August 31, 2026).
• Adaptive thinking on by default with five selectable effort levels (low, medium, high, max, x-high), enabling fine-grained cost-performance tuning.
• Strong reasoning and coding: 91.1% on GPQA Diamond, 39.6% on Humanity's Last Exam, 70.7% on long-context reasoning (AA-LCR).

Where it lags

• Opus 4.8 remains stronger for the deepest agentic reasoning and cybersecurity tasks.
• Uses an updated tokenizer that produces ~30% more tokens for the same text compared to Sonnet 4.6 — effective cost per request can differ even though per-token pricing is unchanged.
• Closed, API-only with no open weights.

Best for: high-volume agentic pipelines, coding workflows, and multi-step agent tasks where Opus 4.8 is cost-prohibitive.

What Claude Sonnet 5 Is

Claude Sonnet 5 is Anthropic's sixth-generation Sonnet model and the first Sonnet designed explicitly for the agentic era. Anthropic describes its design goal as making sustained autonomous execution practical at mid-tier cost: the model finishes complex multi-step tasks where Sonnet 4.6 would stop short, self-checks its own output without being asked, and does so at a price point that makes it deployable at scale.

The biggest architectural change from Sonnet 4.6 is the promotion of adaptive thinking from opt-in to the default mode. Every request now benefits from internal reasoning, with five effort levels letting developers dial cost and latency to match the task. For straightforward tasks, low effort delivers fast, cheap responses; for complex coding or research tasks, x-high effort approaches Opus 4.8's output quality. This makes Sonnet 5 effectively a range of models in one API identifier.

Sonnet 5 also introduces a new tokenizer that processes text into approximately 30% more tokens than Sonnet 4.6. Pricing per token is unchanged ($3/$15, introductory $2/$10), but teams migrating from Sonnet 4.6 should recount prompts and revisit max_tokens budgets before assuming cost parity. Anthropic set introductory pricing to make the transition roughly cost-neutral for typical workloads.

For Benchgen, Sonnet 5 is directly relevant: its variable effort model creates a family of cost-performance operating points that are exactly the kind of nuance production benchmarking surfaces and single-number leaderboards obscure.

Specifications

Pricing

Note: the new tokenizer produces ~30% more tokens for equivalent text vs. Sonnet 4.6, so effective cost per request may differ even though per-token rates are unchanged. Source: Anthropic — Introducing Claude Sonnet 5.

Key Benchmark Results

Source: OpenRouter — Anthropic: Claude Sonnet 5, scores from Artificial Analysis.

Related Models & Benchmarks

• Claude Sonnet 4.6 — predecessor; Sonnet 5 is a drop-in upgrade
• Claude Opus 4.8 — higher-capability sibling; Sonnet 5 narrows but does not close the gap