Claude Sonnet 5: Anthropic's Cheaper, Most Agentic Sonnet Yet

Anthropic has launched Claude Sonnet 5, and it's pitched squarely at people building agents: "our most agentic Sonnet yet," in the company's words — a model that makes plans, drives tools like browsers and terminals, and runs autonomously at a level that, a few months ago, needed a larger and more expensive model. The headline for developers isn't just the capability jump, though. It's the price: Sonnet 5 delivers most of Opus-tier agentic performance at a fraction of the cost, and it's now the default model across free and paid Claude plans.

What's new

Announced June 30 and rolling out July 1, 2026, Sonnet 5 is Anthropic's mid-tier model recast as an agent workhorse. The pitch is autonomy and tool use: planning multi-step tasks, calling browsers and terminals, and staying coherent across long agentic loops without hand-holding. Anthropic has made it the default model for free and Pro plans, and it's available on every subscription tier — a signal that the company expects most everyday and agentic traffic to run here rather than on the flagship.

The real story is pricing

Sonnet 5's launch price is the aggressive part. Through August 31, 2026 it costs $2 per million input tokens and $10 per million output (introductory), after which it moves to $3 input / $15 output. Against Anthropic's flagship Opus 4.8 ($5 input / $25 output), that's roughly 40% cheaper per token at standard pricing and about 60% cheaper during the introductory window. For agent workloads — which burn tokens across many tool-call round trips — that per-token gap compounds fast.

How it performs

On agentic coding, Sonnet 5 scores 63.2%, up meaningfully from Sonnet 4.6's 58.1% and closing much of the gap to Opus 4.8's 69.2%. On a knowledge-work benchmark, Anthropic says Sonnet 5 actually slightly outperforms Opus 4.8. The pattern is clear: Sonnet 5 doesn't beat the flagship on the hardest coding tasks, but it lands close enough that the price difference makes it the rational default for a large share of agent work — and it can exceed the flagship on some non-coding tasks.

Cursor vs Claude Code vs Codex: Which AI Coding Tool Is Worth It?

What it means if you build agents

The practical takeaway is a re-tiering of your model routing. For agent loops, batch jobs, and tool-heavy pipelines where token volume is high and each individual step doesn't demand frontier reasoning, Sonnet 5 is now the obvious default — you get most of the agentic capability at 40–60% less cost. Reserve Opus 4.8 for the hardest, correctness-critical coding and long-horizon reasoning where the extra six points on the agentic-coding benchmark actually change the outcome. Because Sonnet 5 is the platform default, teams on Claude Code and the API will get it unless they pin a model explicitly — worth checking your configs so you're on the tier you intend.

Bottom line

Claude Sonnet 5 is less a raw-capability moonshot than a cost repricing of agentic AI: near-Opus agent performance at roughly half the token price, shipped as the default everywhere. For anyone running agents at scale, the calculus just shifted — the question is no longer "Sonnet or Opus for this task?" but "is this task hard enough to justify paying for Opus at all?" For a growing share of agent work, the honest answer is no.

Sources and further reading