Claude Sonnet 5 as your default agent runner: what I saw in terminal loops
Claude Sonnet 5 handles terminal and browser agent loops well enough to replace Opus for most tasks — but don't build your budget on the intro pricing.
Sonnet 5 holds up as a default agent runner for terminal and browser loops that used to need Opus. It's cheaper than Sonnet 4.6 through Aug 31, and for anything under ~15 tool calls per task it now handles the loop without me babysitting it. The catch: the intro pricing is temporary, so don't rebuild your cost math around numbers that expire.
Anthropic made Claude Sonnet 5 the default for Free and Pro users worldwide starting July 1, and pitched it as its most agentic midsize model. According to the original TechCrunch report, it runs plans, browsers, and terminals autonomously at a level that recently required larger models. I spent a day pointing it at real agent workloads to see where that claim survives contact.
Does it actually run terminal loops?
Yes, mostly. I gave Sonnet 5 a repo with a flaky test suite and told it to fix the failures. It ran `pytest`, read the tracebacks, patched three files, re-ran, and stopped when green. No hand-holding on the loop itself.
Where it wobbled: on a task that needed 20+ sequential shell commands with state carried between them, it lost track of an env var it had set earlier and re-set it wrong. Opus 4.8 didn't do that on the same task. So the "near Opus" framing holds for short-to-medium loops and breaks on long stateful chains.
What about browser automation?
Browser loops were the pleasant surprise. Sonnet 5 navigated a multi-step form flow, handled a modal it wasn't told about, and recovered from one misclick by re-reading the page state. It's not perfect — it still burns tokens describing the DOM to itself — but the recovery behavior is the thing you actually pay Opus for, and Sonnet 5 has most of it.
One concrete gotcha: it's eager to declare success. On two runs it reported the task complete while the final submit button hadn't fired. Add an explicit verification step to your prompt or you'll trust a false green.
Does the intro pricing change your cost math?
Sort of, and this is where I'd be careful. Through Aug 31 it's cheaper than Sonnet 4.6, which is genuinely aggressive for a model this capable. If you're running agents at volume, the per-task cost on a browser loop dropped enough that I'd move default traffic off Opus 4.8 for anything that isn't a long stateful chain.
But intro pricing is a hook. Build your budget on the post-Aug 31 rate, not the promo. If the math only works at intro pricing, you don't have a working cost model — you have a countdown timer. The real question is what Sonnet 5 costs in September against whatever OpenAI and Google ship in the meantime.
Should you use it?
Yes, as your default. Sonnet 5 is the right model to point agent traffic at now: it clears the bar for terminal and browser loops that used to justify Opus, and the price is right even after the promo ends. Keep Opus 4.8 in your back pocket for long stateful chains and tasks where a wrong env var costs real money. That's the split I'm shipping with.
FAQ
Is Sonnet 5 good enough to replace Opus 4.8 for agents?
For short-to-medium terminal and browser loops, yes. For long stateful chains with state carried across 15+ tool calls, Opus 4.8 still tracks context more reliably. Route by task length.
Will the intro pricing last?
No. The below-Sonnet-4.6 pricing runs through Aug 31. Model your budget on the standard rate after that, or you're building on a number that disappears.
What's the biggest gotcha in practice?
Sonnet 5 declares success early. On browser flows it reported tasks complete before the final action fired. Add an explicit verification step to your prompt and don't trust the first green.