Claude Sonnet 5: Release Date, Pricing, and Benchmarks

Claude Sonnet 5 is Anthropic's newest Sonnet-tier model, released on June 30, 2026, priced at $3 per million input tokens and $15 per million output tokens, with an introductory rate of $2/$10 through August 31, 2026. It ships a 1M-token context window, 128K max output, and benchmark scores within a few points of Claude Opus 4.8 at 60% of the price (40% during the intro window). Everything on this page is verified against Anthropic's launch materials and documentation as of July 2026.

What Is Claude Sonnet 5?

Claude Sonnet 5 is Anthropic's mid-tier flagship, positioned as the best combination of intelligence, speed, and cost in the Claude lineup, and Anthropic calls it the most agentic Sonnet yet. The official announcement emphasizes agentic work: planning, using tools like browsers and terminals, and running autonomously on long tasks, with performance Anthropic describes as close to Opus 4.8 at lower prices.

TechCrunch framed the launch as "a cheaper way to run agents," which matches the pricing strategy: an introductory discount aimed directly at agent workloads that meter tokens by the billion. Anthropic also reports lower hallucination and sycophancy rates than Sonnet 4.6, and the ability to finish complex tasks its predecessor could not complete.

Specs and Pricing

Claude Sonnet 5 costs $3 per million input tokens and $15 per million output tokens at standard rates, with introductory pricing of $2/$10 through August 31, 2026. The full spec sheet:

One cost caveat that headline pricing hides: Sonnet 5 uses a new tokenizer that counts roughly 30% more tokens than Sonnet 4.6 on the same text, as noted in Simon Willison's launch analysis. Do not reuse token budgets measured on 4.6-era models, and re-benchmark real costs on your own traffic rather than multiplying old counts by new rates.

Claude Sonnet 5 Benchmarks

On Anthropic's published evaluations, Claude Sonnet 5 scores 63.2% on SWE-bench Pro and 80.4% on Terminal-Bench 2.1. Those two numbers place it a few points behind Opus 4.8 (69.2% and 82.7%) and far ahead of Sonnet 4.6, which scored 67.0% on Terminal-Bench 2.1; the 13-point Terminal-Bench jump in one release is the headline of the launch. Figures are from Anthropic's launch materials, as compiled in Vellum's benchmark breakdown.

Anthropic's migration guidance adds a practical benchmark of its own: Sonnet 5 at medium effort is comparable to Sonnet 4.6 at high effort, and it reaches what was previously Opus-tier quality on many coding tasks. We have not independently benchmarked Sonnet 5, so treat all of the above as vendor-published numbers; they are consistent across Anthropic's announcement and third-party compilations, but they are not independent measurements.

What Changed From Sonnet 4.6

The biggest changes from Sonnet 4.6 are agentic reliability, the new tokenizer, and the removal of sampling parameters. Concretely, as of the June 30, 2026 release:

• Agentic execution: Anthropic reports Sonnet 5 completes complex, long-horizon tasks Sonnet 4.6 could not finish, and checks its own output without being asked.
• Terminal-Bench 2.1: 67.0% to 80.4%, closing most of the gap to Opus in a single release.
• Adaptive thinking: on by default when the thinking parameter is omitted.
• API surface: temperature, top_p, and top_k are no longer supported; effort levels are the tuning knob instead.
• Tokenizer: roughly 30% more tokens counted on the same English text versus 4.6.
• Safety: Anthropic reports better refusal behavior and lower hallucination and sycophancy rates, with safeguards similar to the Opus 4.7 and 4.8 tier, per the Sonnet 5 system card.

Sonnet 5 vs Opus 4.8

Opus 4.8 stays ahead on the hardest work, but Sonnet 5 delivers most of its capability at 60% of the per-token price, and as little as 40% during the introductory window. The gap is about 6 points on SWE-bench Pro and about 2 points on Terminal-Bench 2.1, both models share the 1M context window and 128K output ceiling, and they use the same tokenizer, so cost comparisons are clean.

The working rule we recommend: default to Sonnet 5 for everyday and iterative work, and promote long autonomous runs, sprawling multi-file changes, and expensive-to-retry tasks to Opus 4.8. The full decision framework, including when paying for Opus is a mistake, is in our Claude Opus vs Sonnet comparison. A tier above both, Claude Fable 5 launched June 9, 2026 but has been suspended since mid-June; our Fable 5 and Claude Code explainer covers that situation.

Availability: claude.ai, Claude Code, and the API

Claude Sonnet 5 is available on every Claude surface as of July 2026: it is the default model for Free and Pro users on claude.ai, available to Max, Team, and Enterprise plans, live in the Claude API under claude-sonnet-5, and selectable in Claude Code with the /model command. Model listings are in Anthropic's model documentation.

For subscription users, that default status matters: Free-tier users now get a near-Opus coding model at no cost within usage limits (see is Claude free), and Pro users get it with materially higher limits (see our Claude Pro guide). In Claude Code, Sonnet 5 as the iteration model with Opus 4.8 reserved for the hardest tasks is the cost-effective pattern.

Limitations and Tradeoffs

Three things to keep honest. First, all benchmark numbers here are Anthropic's published figures, not independent tests. Second, the tokenizer change means real-world costs can land higher than the sticker price suggests, especially against budgets calibrated on Sonnet 4.6. Third, the intro pricing expires August 31, 2026, so any cost model built on $2/$10 needs a plan for the 50% jump to standard rates. And if your workload involves genuinely hard, long-horizon autonomous runs, Opus 4.8 still measurably leads; Sonnet 5 narrows the gap but does not close it.

Related Guides

• Claude Opus vs Sonnet: Opus 4.8 vs Sonnet 5 in 2026
• Best Claude Models in 2026, Ranked
• Claude Pro: Price, Limits, and What You Get
• Claude Max: Who Actually Needs It