Anthropic Launches Claude Sonnet 5: Most Agentic Model Yet

Claude Sonnet 5 is out, and Anthropic is positioning it as the most agentic Sonnet-class model it has shipped. The system plans multi-step work, drives tools like browsers and terminals on its own, and pushes tasks through to completion with less hand-holding than earlier Sonnet releases needed. Every Claude plan gets it starting now, and developers can call it through the API under the identifier claude-sonnet-5.

Anthropic built its reputation for agentic skill on the Sonnet 3.5 through 3.7 releases, though the company has acknowledged that its Opus-class systems pulled ahead on that front more recently. Sonnet 5 pushes back the other way. Anthropic puts its overall capability near Opus 4.8's level, at a fraction of the cost, and describes broad gains over Sonnet 4.6 in reasoning, tool use, coding, and everyday knowledge work.

Where Sonnet 5 Actually Matches Opus 4.8

Two specific benchmarks anchor that comparison: BrowseComp, which tests agentic search, and OSWorld-Verified, which tests computer use. Push the effort setting higher on either one, and Sonnet 5's scores can land in Opus 4.8 territory, though Anthropic is careful to note Opus still wins on raw capability. The tradeoff is left to the developer: dial effort up for tougher jobs, dial it down to save on cost.

Daniel Shepard, a senior engineer who tested Sonnet 5 ahead of launch, gave it a two-step assignment: update Salesforce account tiers, then email enterprise contacts about the release. He watched it complete both parts without pausing in between. Earlier Sonnet versions, he said, usually stalled halfway through.

What It Costs and Who Gets It First

Rollout starts immediately: Claude Sonnet 5 becomes the default model for Free and Pro users, while Max, Team, and Enterprise plans get access alongside Claude Code and the Claude Platform.

Anthropic is running introductory pricing through August 31, 2026, at $2 per million input tokens and $10 per million output tokens; standard rates of $3 and $15 take over after that. Opus 4.8, by comparison, costs $5 and $25 for the same units, which puts real distance between the two models on price alone.

One detail changes the math. Sonnet 5 runs on an updated tokenizer, so the same prompt can consume up to 1.35 times more tokens than before, depending on what's in it. Anthropic says the introductory rate was set specifically to offset that shift and keep the move from Sonnet 4.6 close to cost-neutral. Rate limits also went up across Chat, Cowork, Claude Code, and the Platform to handle the extra load from higher effort settings.

The Safety Tradeoffs Anthropic Is Disclosing

Pre-deployment testing put Sonnet 5 ahead of Sonnet 4.6 on several safety measures, according to Anthropic: stronger resistance to prompt injection attempts, and lower rates of both hallucination and sycophancy. Cybersecurity is the exception. The model's ability to carry out cyber tasks stays well under what Anthropic's Opus-tier systems can do, and the company ships it with the same real-time cyber safeguards already active on Opus 4.7 and 4.8.

Free-tier users can now turn on autonomous browser and terminal control, a capability that used to sit behind paid tiers. The real test is how fast teams fold Sonnet 5 into daily workflows before the August 31 pricing window closes. Anthropic lays out the full benchmark data and safety findings in its Claude Sonnet 5 announcement.