Sign in Subscribe
news

Claude Opus 5.5 - Fable-level work for less than half the price

Sebastien Dubois

Sebastien Dubois

— 3 min read

Canonical version: Claude Opus 5.5 - Fable-level work for less than half the price.

Less than two months after Claude Opus 5, Anthropic just released Claude Opus 5.5. It's the first model of the new Claude 5.5 family, and Sonnet 5.5 and Haiku 5.5 are coming "in the coming weeks".

According to Anthropic, it performs at the level of Claude Fable 5.1 on most tasks, and costs 40% less to run than Opus 5. Fable 5.1 costs $10/$50 per million tokens. Opus 5.5 costs $4/$20, 60% less per token.

In this piece, I'll go over what changed for those of us who use Claude every day, what breaks if you build on the API, and how its safeguards route some requests to older models.

Cheaper, and not just per token

Token prices dropped 20% ($4 input, $20 output), and cache reads dropped 60% to $0.20. The "40% cheaper" figure comes from efficiency: the model uses fewer tokens and fewer steps to finish the job, and it's more than 30% faster.

Customers report similar results. Box saw about a third of the tokens Opus 5 used. GitHub says it solved more terminal tasks than Opus 5 in VS Code in less than half the steps. Deloitte got 72% of known bugs caught in code review at the LOWEST effort setting, against 56% for Opus 5 at high effort. Those are internal evals, not independent benchmarks, so take them with a grain of salt. Still, several companies report the same pattern.

On Anthropic's benchmarks, it beats Fable 5.1 almost everywhere (66.4% vs 55.8% on Terminal-Bench 4.0, for instance). Anthropic itself says not to read too much into those margins, because a few points don't make a noticeable difference in real work at this level. GPT-6 Astra still wins on agentic science and AutomationBench, but Opus 5.5 beats it on FrontierCode at roughly 20% of the per-task cost.

It finally writes clearly

A common complaint about Opus 5 was dense, hard-to-follow output. Opus 5.5 puts the important information first, uses less jargon, and follows the writing rules you give it. Jackson Kernion, who worked on it, says he hasn't been this happy with a model's writing since Opus 4.6.

Dan Shipper's tests at Every back this up partially: the highest reading ease score they've measured, but it still buries the main point in intro paragraphs. I read Claude's output for hours every day in Claude Code, so clearer writing matters a lot to me.

What breaks on the API

If you use the API directly, don't just swap the model ID:

  • Thinking can't be disabled anymore. Use the effort parameter instead (the default is now medium, it was high on Opus 5)
  • Forced tool use (tool_choice of type any or tool) returns an error
  • Thinking blocks are tied to the model and the conversation. Keep your conversations append-only
  • The old computer_20251124 tool is rejected on the Claude API and Google Cloud

Also, the text the model writes between tool calls now comes back in thinking blocks, empty by default. If your UI shows progress updates, it will stop showing them, without any error.

Safeguards route some requests to older models

Opus 5.5 is comparable to Mythos 5.1 in biology and cybersecurity, so it ships with Fable-class safeguards. When a classifier fires, your request is answered by an OLDER model: Claude Opus 4.8 for most cybersecurity requests, Opus 5 for biology and frontier LLM work.

In a multi-step agent run, that means some steps may be handled by a different, weaker model than the one you selected. Your evals probably assume one model per run. Handle refusals explicitly, and log which model actually answered.

On alignment, Anthropic reports its best score to date, with METR and Frontier Design testing before release. It's also the first model since Dario Amodei called for "pacing the frontier". One caveat Anthropic admits: the model often suspects it's being tested, which makes those results harder to read.

What I'd do

  • In Claude Code, Opus 5.5 is already the default for Pro, Max and Team plans, at medium effort. Rate limits go about 25% further than with Opus 5, and the five-hour limits went up. Try medium before reaching for high
  • Re-read your CLAUDE.md. A model that follows writing rules more closely will also expose the vague or contradictory ones
  • If you pay for Fable 5.1, run your own evals. For most work, Opus 5.5 looks like the better deal
  • If you're cost-sensitive, wait a few weeks for Sonnet 5.5 and Haiku 5.5

The full breakdown (benchmarks table, API details, reception) is in my Claude Opus 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