Claude Sonnet 5.5
Anthropic's mid-tier model: two points behind Opus 5.5 on a test of real-world work, at half the price. I still use it as a subagent rather than pick it by hand.
Quality
Modality
multimodal
Context
1M tokens
Access
closed
Fabian's Take
"A curious release: almost as good as Opus, and cheaper. I have a feeling I'll still stick with Opus in my existing setup and let it hand the smaller, clearly defined jobs to Sonnet as a subagent."
Claude Sonnet 5.5 replaces Sonnet 5 as Anthropic’s mid-tier model, released on 28 September 2026, six days after Opus 5.5. It’s the second model in the Claude 5.5 family. Haiku 5.5 is still announced for “the coming weeks.”
What changed
The price didn’t: $2 per million input tokens and $10 per million output, with cache reads at $0.20 per million. What you pay per task still goes down. By Anthropic’s numbers, Sonnet 5.5 gets through the same work with fewer tokens, so most work costs up to 30% less, and its output arrives more than 30% faster than Sonnet 5’s.
The bigger change is how close it now sits to Opus. On Artificial Analysis’s GDPval-AA, a test of real-world work across 44 occupations, Sonnet 5.5 scores 1844. Sonnet 5 scored 1449, and Opus 5.5 scores 1846. On Terminal-Bench 4.0, which tests agents doing work on the command line, it reaches 70.6%. Anthropic positions it for well-scoped everyday tasks, fixing bugs, and producing polished documents, slides and spreadsheets.
What to know before you migrate
If your code switched thinking off, move it to the between_tools thinking setting before upgrading, because turning thinking off is no longer accepted. If you work through the Claude apps or Claude Code rather than the API, this doesn’t touch you.
Where it sits in the lineup
On paper, the gap to Opus is now tiny, and Sonnet costs half as much per token. I’ll still stick with Opus 5.5 in my existing setup. Opus plans and decides, then hands the smaller, clearly defined jobs to Sonnet as a subagent: searching a codebase, drafting a section, running a well-defined change. That’s where the lower price pays off, because those jobs burn a lot of tokens without needing the top model.
If you’re building on the API and paying per token for volume work, this is the Claude model to test first. If you work in the Claude app, try it on your next well-defined task and see whether you notice the difference from Opus. For more on matching the model to the job, see Think Expensive, Execute Cheap.
The Verdict
Best for: Well-scoped everyday jobs, bug fixes, polished documents and slides, and the focused jobs a stronger model hands to a subagent.
Pros
- Two points behind Opus 5.5 on Artificial Analysis's GDPval-AA test of real-world work (1844 against 1846), at half Opus's per-token price
- More than 30% faster output than Sonnet 5, and up to 30% cheaper per task at the same token price
- 70.6% on Terminal-Bench 4.0, a test of agent work on the command line
Cons
- Code that switched thinking off has to move to the between-tools setting before upgrading
- On the hardest, most open-ended work, Opus is still the safer choice
Specs
- Pricing $2/M input, $10/M output · cache read $0.20/M, cache write $2.50/M
- Cost Tier moderate
- ⚡Speed Tier fast
- License Proprietary
