Claude Sonnet 5 ships as default for Free and Pro, undercuts Opus on agentic tasks

Anthropic's mid-tier model gets a two-month introductory price of $2/$10 per million tokens and a new tokenizer that reshapes cost math.

Anthropic shipped Claude Sonnet 5 on Tuesday and immediately installed it as the default model on Claude Free and Pro, with availability across Max, Team, and Enterprise. The framing is pricing, not capability: $2 per million input tokens and $10 per million output through August 31, 2026, after which the rate reverts to Sonnet 4.6’s $3/$15.

The competitive geometry is legible. Sonnet 5 undercuts Opus 4.8, OpenAI’s GPT-5.5, and Google’s Gemini 3.1 Pro on sticker price, while Gemini 3.5 Flash remains cheaper still. The pitch is that mid-tier is now good enough to run agents against, which is where inference spend actually accumulates.

The benchmarks tell the intended story. On Anthropic’s agentic coding eval, Sonnet 5 lands at 63.2%, above Sonnet 4.6’s 58.1% and beneath Opus 4.8’s 69.2%. On Terminal-bench 2.1, TechRadar reports Sonnet 5 at 80.5% against Sonnet 4.6’s 67%. TechCrunch notes it slightly outperforms Opus 4.8 on a knowledge-work benchmark, which is the kind of result that makes the mid-tier default sensible.

Read the fine print, though. Anthropic says the new tokenizer produces roughly 30% more tokens for the same input text, which quietly reshapes per-request cost math even at the promotional rate. Adaptive thinking is on by default; manually invoking extended thinking returns HTTP 400, as do non-default temperature, top_p, and top_k. It’s the first Sonnet with real-time cybersecurity safeguards, and refused requests come back as HTTP 200 with stop_reason: refusal. Anthropic reports “much lower” cybersecurity-task capability than current Opus models.

Context is 1M tokens by default, 128k max output. Fable 5 returns globally July 1. Anthropic is also proposing an industry jailbreak-severity scoring framework with Amazon, Microsoft, Google, and other Glasswing partners.

The backdrop is a confidential SEC filing on June 1. A cheaper default model, benchmark parity claims against Opus, and a coordinated safety framework are exactly the artifacts a company assembles before it prices an IPO.

Claude Sonnet 5 ships as default for Free and Pro, undercuts Opus on agentic tasks

Sources