2026-06-30 — Tuesday
Anthropic released Claude Sonnet 5 today, a more agentic version of its midsize model, pitched squarely at the cost of running agents rather than at the top of the benchmark tables. TechCrunch framed the launch as “a cheaper way to run agents”, and that is the honest summary: the interesting number here is not the peak score, it is what the peak score costs.
Anthropic calls it the most agentic Sonnet model yet, with the gains concentrated in reasoning, tool use, coding and knowledge work. The practical claim is that Sonnet 5 can plan, drive a browser or a terminal, and run without supervision at a level that a few months earlier needed a larger and more expensive model.
Anthropic also reports better self-checking and follow-through on long tasks, and lower hallucination and sycophancy rates than its predecessor.
Sonnet 5 lands at $2 per million input tokens and $10 per million output, described as permanent pricing rather than an introductory rate. Set against Opus 4.8 at $5/$25, that is the whole argument: benchmark data shows Sonnet 5 closing the gap with the larger model at a fraction of the cost.
For anyone running agents in a loop — where the token bill is a function of how many turns the task takes, not how impressive the demo is — that ratio matters more than a leaderboard position.
The model shipped immediately across Free, Pro, Max, Team and Enterprise plans, in Claude Code, and on the Claude Platform. On safety, Anthropic reports lower rates of undesirable behaviour than Sonnet 4.6 and substantially reduced capability on cybersecurity tasks compared with the Opus models.