Skip to content

AI

Claude Sonnet 5 does agent work without the big-model bill

Anthropic's new midsize model plans, drives browsers and terminals, and works on its own. The pitch: capability that used to need a bigger model, at a lower price tier.

Cover photo: Axisadman / CC BY-SA 3.0

TL;DR

  1. Anthropic ships Claude Sonnet 5, a midsize model built for agentic work.
  2. It plans, uses browsers and terminals, and runs solo, per Anthropic.
  3. OpenAI and Google pitched the same thing weeks ago. Cost is the new fight.

What happened

Anthropic released Claude Sonnet 5, an upgraded version of its midsize model. The headline feature is agentic behavior: it can draft a plan, operate tools such as browsers and terminals, and carry on without a human hovering over it.

Anthropic's own claim, from its blog post, is that this level of autonomy needed larger and pricier models only a few months ago. That is a vendor statement, so treat it as one. The point of Sonnet 5 is to bring that skill set down a tier.

The timing is not subtle. OpenAI launched GPT-5.6 Sol in preview last week, and it is billed as the company's most agentic model yet, letting users split work across subagents for longer autonomous runs. Google's Gemini 3.5 Flash arrived in May with a pitch to move from chatbot to a tool that plans and builds.

Three labs, one script.

Why it matters

When every major lab ships agent skills, agent skills stop being the headline. They become the entry fee. The question shifts from who can do agentic work at all to who does it cheaply and reliably, with fewer humans checking the homework.

That is why the midsize slot matters. A model that handles planning and tool use at a lower price tier is the one companies actually run all day, in loops, at scale. The flagship gets the demo. The midsize model gets the workload.

Agentic is the new baseline. Price and reliability are the new fight.

One caution, sapiens: "close to the big model for less" is a promise until you run it on your own tasks. The pitch is credible, and it is still a pitch.

Mid-size models are where most real work happens. The biggest model gets the headlines, but companies running thousands of agent tasks a day care about price per task and speed.

A cheaper model that is good enough for most steps, with the flagship called in only for the hard parts, is how many teams now build. Sonnet 5 is Anthropic's bid to be that default, and rivals have similar offers on the table.

What to do with it

  • Test it on one real workflow.Pick a boring multi-step task you already do and see whether Sonnet 5 finishes it without hand-holding.
  • Compare cost, not just quality.Run the same task on a larger model and on Sonnet 5, then check what each run costs you.
  • Keep a human checkpoint.Reliability without oversight is the unproven part, so review outputs before anything touches production.
  • Watch the three-way race.GPT-5.6 Sol and Gemini 3.5 Flash chase the same job, so avoid locking your stack to one vendor yet.

Keep up, sapiens. The midsize models just got ambitious.