Anthropic's Claude Sonnet 5 lands as the cheaper default for running AI agents

Anthropic released Claude Sonnet 5 on June 30, 2026, and the pitch is straightforward: most of the agentic capability of a top-tier model, at a fraction of the cost. For anyone learning to build software with AI, it is the kind of release that quietly changes what you can afford to leave running.

What shipped

Sonnet 5 is now the default model for Free and Pro plans, and is available to Max, Team, and Enterprise users. Developers can call it through the Claude API as claude-sonnet-5, and it is the new default inside Claude Code, where it runs with a native 1M-token context window.

Anthropic frames it as "the most agentic Sonnet model yet" — 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." In plain terms: the model that watches your terminal, edits files, and iterates on its own errors no longer has to be your most expensive one.

The pricing is the headline

During an introductory period through August 31, 2026, Sonnet 5 costs $2 per million input tokens and $10 per million output tokens. From September 1 it reverts to $3 / $15 per million — matching the old Sonnet 4.6 price. TechCrunch notes that positions Sonnet 5 below Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on cost (though still pricier than Gemini 3.5 Flash).

Because AI agents burn tokens continuously — reading files, planning, running tests, retrying — the per-token price is often the real ceiling on how much autonomous work you can run. A cheaper capable model means longer sessions and more parallel agents for the same budget.

How capable, actually

On the agentic-coding benchmark TechCrunch reported, Sonnet 5 scored 63.2%, between its predecessor Sonnet 4.6 (58.1%) and Opus 4.8 (69.2%). Anthropic says it "slightly outperforms Opus 4.8" on some knowledge-work tasks that require subtle judgment and deep research. The practical takeaway: Sonnet 5 is a strong default for most coding work, with Opus 4.8 still worth reaching for on the hardest problems.

Anthropic also reports reduced "undesirable behaviors" — lower hallucination rates and better refusal of malicious requests — versus Sonnet 4.6.

The context that matters

The launch coincided with the U.S. lifting export controls on Anthropic's Fable 5 and Mythos 5 models, and Axios framed Sonnet 5 as a bid to win everyday agent workloads on price rather than raw ceiling.

If you build with Claude Code, you already have Sonnet 5 — it became the default in version 2.1.197. Worth knowing before you start a long agent run: you can dial effort levels down on simpler tasks to save tokens, and reach for higher effort (or Opus) when a problem genuinely needs the extra reasoning.