Claude Fable 5.1 - cheaper caching, calmer safeguards, same old Claudish

Canonical version: Claude Fable 5.1 - cheaper caching, calmer safeguards, same old Claudish.

Three months after the launch (and the ban, and the un-ban) of Claude Fable 5, Anthropic just shipped Claude Fable 5.1 and Mythos 5.1.

Same story as last time. Fable and Mythos are one model with two safeguard levels. Fable is for everyone, Mythos is for vetted security and life sciences organizations, US only for now. If you want the full breakdown of that structure, I wrote about it in Claude 5.

In this piece, I want to share what actually changed for those of us using these models every day, and why the community reaction had almost nothing to do with the announcement.

What changed

The capability numbers are real this time. Terminal-Bench-Science went from 24.7% to 52.6%. AutomationBench from 17.1% to 31.4%. And on Terminal-Bench 4.0, Fable 5 was actually behind Claude Opus 5 (42.0% vs 52.3%). 5.1 is now ahead at 55.8%. That little detail explains a lot. Many people paid double for Fable 5 and did not feel the difference on day-to-day coding. Now the ordering makes sense again.

The showcase is science: protein binders with 10x higher affinity than competition entries, a Venus elevation map at 2-3km resolution instead of 10-20km, deep learning models sped up by 2.5x. Impressive, but not what most of us will use it for.

What most of us WILL notice:

• Cache reads dropped to $0.25 per million tokens, down 75%. Input and output prices did not move ($10/$50). Anthropic estimates 25% savings on typical workloads and up to 45% on agentic ones. Agentic loops re-read the same context over and over, so this is where the money actually goes
• Safeguards produce 60% fewer false positives in the cybersecurity domain. Fable 5 was so trigger happy that Anthropic itself admitted biology was "practically unusable". Good riddance
• It goes deeper before it patches. Anthropic claims it finds root causes rather than applying shortcuts. Millennium says it found a rare crash "nobody on our team had explained in four to five years". MongoDB built a "complex prototype in about three days", mostly unattended
• Anti-distillation is visible now. New accounts cannot edit prior context. Remember the "secret sabotage" mess with 5.0, where the model silently degraded answers it suspected were distillation attempts? Lesson learned, apparently

Every summed it up as "Fable-level intelligence, Opus-level price, Sonnet-speed". I would not go that far on the price, but the cache cut does change the math.

What the community talked about instead

Now the fun part. The Hacker News thread barely mentioned benchmarks.

An Anthropic engineer opened by praising the "much more natural style" of 5.1. The thread answered with pages of complaints about what people now call "Claudish": dense, jargon-inventing, hedge-stacking prose that sounds confident while saying very little. One commenter's verdict stuck with me: "It's both dense and vacuous. Dense because it's full of jargon it's made up, and vacuous because even with all that it's not actually saying much."

The recurring complaints:

• You can tell it to write plainly in a system prompt or a CLAUDE.md, and it forgets within a few turns. "Training supersedes random markdown files"
• Code comments that reference files that don't exist and dates that were never real. The model writing notes for itself instead of for humans
• Workarounds people actually use: pipe the output through Haiku, use a competitor to clean it up, demand "Simplified Technical English", ban superlatives explicitly

I've been there too. I maintain a whole humanizer skill in my vault for exactly this reason, and I still catch new tics every month. The model gets smarter with each release. The writing habit does not go away. It is baked into the training, not into the prompt, and no point release will fix that for you.

My take

Two things.

First, point releases in this generation are where the usability fixes land. Cheaper caching, calmer safeguards, longer autonomy. The big version numbers are for capability. If you skipped Fable 5 because of cost or because it refused half your security questions, 5.1 is the one to re-evaluate.

Second, and this matters more to me: the Hacker News thread is a reminder that we are still responsible for what we ship, text included. The model can now design proteins. It still cannot be trusted to write a paragraph you would sign with your own name without a pass. As I argued in Mastering concepts matters more than ever, the last mile is ours. That's just a fact.

Plan for a humanizing pass, not a prompt fix.

That's it for today! ✨

References

• Anthropic announcement: https://www.anthropic.com/claude-fable-and-mythos-5-1
• Hacker News discussion: https://news.ycombinator.com/item?id=49525378

Related

• Claude Fable 5.1
• Claude Fable 5
• Claude 5
• Claude Opus 5
• Anthropic
• Claude Code Prompt Caching
• Quality non-fiction is the antithesis of AI slop
• Fable 5 ban

About Sébastien

Ready to get to the next level?