Sign in Subscribe
news

Claude Sonnet 5.5 - Opus-level everyday work at half the price

Sebastien Dubois

Sebastien Dubois

— 3 min read

Canonical version: Claude Sonnet 5.5 - Opus-level everyday work at half the price.

Six days after Claude Opus 5.5, Anthropic released Claude Sonnet 5.5, the second model of the Claude 5.5 family. Haiku 5.5 is next, "in the coming weeks".

It costs the same as Claude Sonnet 5 ($2 input, $10 output per million tokens), 30%+ faster, and up to 30% cheaper per task because it needs way fewer tokens. On several benchmarks it lands within two or three points of Opus 5.5, which costs twice as much per token.

It's also now the model behind the free tier on claude.ai. As Simon Willison pointed out, that makes Claude's free offering a LOT more capable than ChatGPT's.

Benchmarks and efficiency

Terminal-Bench 4.0 (agentic coding in a terminal) went from 10.3% on Sonnet 5 to 70.6%, which is above Opus 5.5 (66.4%). On GDPval-AA (real-world knowledge work), it's 2 Elo points behind Opus 5.5 and about 400 above Sonnet 5. Artificial Analysis ranks it #2 on its Intelligence Index, right behind Opus 5.5.

Balyasny measured about 121k tokens per answer vs 497k for Sonnet 5, with a better score. Base44 got Opus 5-level apps in 3.6 iterations instead of 7.7. CodeRabbit's reviews cost about 60% less and take half the time. These are internal evals, and all three report fewer tokens, fewer iterations or a lower cost.

Don't max the effort dial

At low and medium effort, Sonnet 5.5 is cheap for what you get: Anthropic says it beats Sonnet 5's best score on several benchmarks at about a tenth of the cost.

At max effort, it's the opposite. Artificial Analysis measured $7.60 per task at Max, with about 193k output tokens per task (a record). Opus 5.5 at Max costs $5.98 per task and scores higher. At High, Sonnet 5.5 costs $1.08 per task, which is where it's actually competitive.

More effort can even make results worse: on FrontierCode, Max scores 46.2% vs 52.1% at Xhigh, because the model started spinning up review subagents, timing out, and making edits nobody asked for. And Simon Willison's pelican test at Max burned 128k thinking tokens without producing anything.

If a task needs Max effort, I'd use Opus 5.5 instead.

It's not an Opus replacement

Anthropic says so itself: Opus 5.5 "remains clearly stronger at complex, open-ended work requiring sustained judgment". CodeRabbit's code review tests agree: on hard real-world PRs, Sonnet 5.5 caught 6 of 13 known issues, vs 8 for Opus 5.5, with much lower precision.

Sonnet 5.5 is for well-scoped work: bug fixes, fast iteration on features, documents, slides, spreadsheets, and repeated agent tasks. Testers also say it has a good eye for design.

What breaks on the API

If you're migrating from Sonnet 5, don't just swap the model ID:

  • thinking: {"type": "disabled"} returns an error. Use the new between_tools setting instead (only at low, medium or high effort)
  • Forced tool use (tool_choice of any or tool) is gone. Use auto with strict tools
  • Responses now contain thinking blocks by default, so code that reads content[0].text breaks
  • Keep conversations append-only: thinking blocks are tied to the model and the account

On the safeguards side, it's the first Sonnet with Opus-style cyber safeguards: high-risk cybersecurity requests visibly fall back to Sonnet 5. Handle refusals explicitly, and log which model actually answered.

My recommendations

I'd use Sonnet 5.5 at medium for well-scoped work. That's the default in Claude Code and the Claude apps, and it's where the model is the best value. For harder Sonnet tasks, go to high; if a task seems to need max, switch to Opus 5.5 instead.

Plan with Opus, execute with Sonnet. Let Opus 5.5 set the architecture and make the hard calls, then hand implementation to Sonnet 5.5 subagents, which is cheaper and faster.

  • In Claude Code, the sonnet alias now points to Sonnet 5.5 (v2.1.284+). The default model is still Opus 5.5
  • On the API, remove your old "don't be lazy" prompts and re-run your evals. Those workarounds were written for a different model
  • Track cost per outcome rather than cost per token: at the same list price, the bill varies a lot with the effort level

The full breakdown (benchmarks, effort/cost table, API details, safeguards, reception) is in my Claude Sonnet 5.5 note.

That's it for today! ✨

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe