Agentic Coding Just Became the Cheap Default: What Claude Sonnet 5 Means for Your Team This Week
A Sonnet-class model that plans, uses a terminal and a browser, and finishes multi-step tasks — now the default, at $2/$10 per million tokens. Here's how a small team puts it to work this week.
A Sonnet-class model that plans, uses a terminal and a browser, and finishes multi-step tasks — now the default, at $2/$10 per million tokens. Here's how a small team puts it to work this week. What happened On June 30, 2026 , Anthropic released Claude Sonnet 5 , which it calls its most agentic Sonnet to date. (An early aggregator write-up dated it "July 15" — the correct date is June 30.) The important part isn't a raw-quality leap. It's that Opus-grade agency — planning, tool use, and multi-step follow-through — now ships in a cheaper Sonnet-class model: - It is the default model for Free and Pro plans, and it's available on Max, Team, Enterprise, Claude Code, and the API . - It can plan a task, use tools (a terminal, a browser), read the results, and keep working through long multi-step chains — finishing tasks where earlier versions would stall halfway. - It's strongest exactly where real teams hurt: brownfield code — race conditions, hidden tests, tracing a failure to its root cause and shipping a durable fix. - Introductory API pricing is $2 per million input tokens and $10 per million output tokens , running through August 31, 2026 . Why it matters This is the most immediately leverageable release of the week, for one reason: the strongest agentic Sonnet is now the model your team already uses by default , priced cheaply enough to run over real workloads. "Finishes the job" is the whole story. The gap between an AI demo and an AI workflow is follow-through. A model that stops halfway through a multi-step task needs a human to close every loop. A model that plans, uses tools, and keeps going until the task is done can own a defined chore end to end. That's the difference between "AI drafts a first pass" and "AI closed this ticket." The economics changed. At $2/$10 per million tokens, letting an agent iterate — run the tests, read the failure, try again — is no longer a budget conversation. You can afford to put an agent on real, repetitive engineering work at scale. It's brownfield, not greenfield. The benchmark that matters here is 63.2% on SWE-bench Pro (resolving real issues in real repositories), alongside 81.2% on OSWorld-Verified (operating a real desktop/OS environment) and 57.4% on HLE (broad, deliberately hard reasoning). It beats its predecessor on every published benchmark. Debugging legacy systems is where a small engineering team is most capacity-constrained — and that's where this lands. Under the hood "Agentic" here means the model runs inside a loop rather than answering a single prompt. It plans an approach, calls a tool (runs a terminal command, fetches a page in a browser, edits a file), reads the result , and decides the next step — repeating until the task is done or it hits a stop condition. Anthropic's framing of the improvement is follow-through: pushing further through long chains before stopping. Distribution is as much the story as the model. The cheapest, most agentic Sonnet is wired straight into the tools peopl