Anthropic released Claude Sonnet 5 on July 1, setting it as the default model for Free and Pro users. The API standard pricing is $3 per million input tokens and $15 per million output tokens, with a promotional period before August 31 at $2 input and $10 output; this is about 60% cheaper than the flagship Opus 4.8 standard pricing.
Claude Sonnet 5 API Pricing, Discount Period, and Tokenizer Details
Claude Sonnet 5 has a context window limit of 1 million tokens, with a maximum output of 128k tokens. This version uses an updated tokenizer. Anthropic explains that the same input under the new tokenizer produces 1.0 to 1.35 times the number of tokens compared to the previous version, depending on content. Anthropic claims the pricing has been adjusted to be "roughly cost-neutral" and recommends that high-traffic users run their own benchmarks to assess actual billing impact.
6 Benchmark Results Published by Anthropic
All data below are from Anthropic, with no third-party independent verification reports yet.
SWE-bench Pro (agentic code): Sonnet 5 63.2%, Opus 4.8 69.2%, Sonnet 4.6 58.1%
Terminal-Bench 2.1 (terminal operation): Sonnet 5 80.4%, Opus 4.8 82.7%
Humanity's Last Exam (multi-discipline reasoning with tools): Sonnet 5 57.4%, Opus 4.8 57.9%
GDPval-AA v2 (knowledge work): Sonnet 5 score 1,618; Opus 4.8 score 1,615
OSWorld-Verified (computer operation): Sonnet 5 81.2%, Sonnet 4.6 78.5%
Firefox 147 Vulnerability Assessment: Sonnet 5 usable vulnerability output rate 0%, partial success rate 13.2%; Sonnet 4.6 partial success rate 8.8%; Opus 4.8 68.8%
Safety Metrics for Sonnet 5 in Anthropic's Safety Report
Anthropic's safety report states that Claude Sonnet 5 has lower hallucination and sycophancy than Sonnet 4.6, stronger ability to reject malicious requests, and this version enables network security protection mechanisms by default. The report also notes that Sonnet 5's rate of inappropriate behavior is still higher than Opus 4.8 and also higher than the strictly restricted Claude Mythos Preview.
Frequently Asked Questions
When was Claude Sonnet 5 set as the default model for Free and Pro users?
Anthropic officially released Claude Sonnet 5 on July 1, 2026, and set it as the default model for Free and Pro users on the same day.
Are the benchmark results published by Anthropic verified by a third party?
All Claude Sonnet 5 benchmark results published by Anthropic are official self-evaluated data. Anthropic has not provided any third-party independent verification reports, nor has it disclosed a timeline for submitting to third-party verification.
What impact does the new tokenizer have on API user billing?
Anthropic explains that the same input under the new tokenizer produces 1.0 to 1.35 times the number of tokens compared to the previous version, and the official claims pricing has been adjusted to be "roughly cost-neutral". Anthropic recommends that high-traffic users conduct their own benchmarks to assess actual billing impact, and does not guarantee that bills will not increase in all scenarios.
