Claude Sonnet 5.5: a cheaper, faster middle model

What actually shipped

Sonnet 5.5 is the second model in Anthropic’s Claude 5.5 family, following Opus 5.5, which we covered when its price cuts landed. The Decoder reports a jump on Terminal-Bench, a coding-focused benchmark, from 10.3 to 70.6 percent. That’s a big swing for what Anthropic still positions as its mid-range model, not its flagship.

TechCrunch reports that Anthropic itself is calling this a “significantly cheaper, faster work partner.” Nobody’s claiming Sonnet 5.5 beats Opus 5.5 outright. The framing backs that up. The implied pitch is that it gets close enough on knowledge-work tasks that the gap stops mattering for a lot of everyday use, while costing less and returning answers faster.

GitHub’s own changelog backs that framing up. It describes Sonnet 5.5 as built for “well-scoped everyday work like building features and fixing bugs,” which is a precise way of saying this isn’t the model you reach for on your hardest architectural problem. It’s the one you reach for on the twenty routine tickets you clear in a normal week, and that’s most of what a solo builder’s coding agent actually does, day to day.

The Decoder also notes that Anthropic has announced Haiku 5.5, which would give it a direct counterpart to each of OpenAI’s three GPT-6 models. That model isn’t out yet, so the lineup isn’t complete. It explains the timing, though: Anthropic is filling out a full price ladder, not just shipping a single upgrade.

Why the benchmark jump matters less than the price cut

A jump from 10.3 to 70.6 percent on Terminal-Bench sounds like the story. It probably isn’t, for most readers of this site. Terminal-Bench measures an agent’s ability to complete tasks in a command-line environment, which correlates with coding-agent usefulness but isn’t a number you’ll feel directly unless you’re running Sonnet inside an agentic coding tool.

What you will feel, if you’re already paying for Claude access through an API or through a tool like GitHub Copilot, is the cost line. Up to 30 percent less per task, on a model that was already the practical choice for most day-to-day coding work, compounds fast if you’re running an agent for hours at a stretch.

If you’ve been tracking intent debt in your own agentic workflows, a faster, cheaper mid-tier model doesn’t fix that problem.

It does lower the cost of the trial-and-error loop that produces it in the first place, which is worth something even if it isn’t a fix. Treat the benchmark numbers as evidence that the smaller model closed a real capability gap, and treat the pricing numbers as the part that shows up on your invoice.

What this means for you

If you’re already using Claude Sonnet through GitHub Copilot, you don’t need to do anything. The upgrade is already live and it’s a straightforward better deal: faster responses, lower per-task cost, similar or better output on the kind of well-scoped work Copilot is built around.

If you’re picking a model fresh, or you default to Opus 5.5 for routine work out of habit, this is a good moment to check whether Sonnet 5.5 covers what you actually need. Opus still makes sense for the hardest problems: a gnarly refactor, a genuinely ambiguous spec, a bug that’s resisted three other models. Everything below that bar is now cheaper to hand to Sonnet, and the gap between the two has visibly narrowed.

This isn’t a release that changes how you build. It’s a release that changes what building costs.

Worth acting on if you’re running an agent for hours a day. Not worth switching tools over if you’re not.