Anthropic ships Claude Sonnet 5 with near-Opus agent performance at $2 per million input tokens

The new mid-tier model, released June 30 and now the default for Free and Pro users, matches Opus 4.8 on a knowledge-work benchmark and closes the gap on agentic coding — at introductory API pricing 60% below the flagship.

Anthropic released Claude Sonnet 5 on Tuesday at $2 per million input tokens and $10 per million output tokens, introductory API pricing that undercuts its own Opus 4.8 flagship by roughly 60% while landing within striking distance of it on the benchmarks that matter for agent workloads. The model is already the default on Claude Free and Claude Pro, and is available across Max, Team, Enterprise, Claude Code, and the Claude API from launch.

The pricing gap is the story. Opus 4.8 lists at $5 input / $25 output; Sonnet 5’s promotional rates hold until August 31, 2026, after which they step to $3 and $15. On GDPval-AA v2, Anthropic’s knowledge-work benchmark, Sonnet 5 scores 1,618 to Opus 4.8’s 1,615. On SWE-bench Pro it hits 63.2% against Opus’s 69.2%, up from 58.1% for Sonnet 4.6. Terminal-Bench 2.1 shows a similar compression: 80.4% versus 82.7%. On OSWorld-Verified, the computer-use eval, Sonnet 5 posts 81.2%.

Anthropic describes the model as able to “make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models.” That’s the pitch to anyone paying per-token for long agent traces, where Opus economics have been the limiting factor.

The system card is more restrained. Sonnet 5 “does not advance our capability frontier,” Anthropic writes, reporting “near-zero” successful stealth rates on autonomy evaluations and “a much lower ability to perform cybersecurity tasks” than current Opus models, which the company attributes to not deliberately training on that domain.

There’s also an updated tokenizer, which maps text to 1.0–1.35× more tokens than Sonnet 4.6 depending on content type. Anthropic calls the introductory pricing “roughly cost-neutral” once that expansion is factored in, meaning the headline discount is smaller in practice than the sticker suggests.

Around all of this sits a quieter Axios detail: Anthropic remains in discussions with the Trump administration over its model lineup, and is still awaiting approval to restore full access to Mythos and Fable 5 after the government asked the company to pull them over security concerns. Mythos is available on a limited basis. The company shipping its cheapest capable agent model while its top-shelf systems sit in regulatory limbo is its own kind of positioning.

Anthropic ships Claude Sonnet 5 with near-Opus agent performance at $2 per million input tokens

Sources