Claude Fable 5.2 Is Out. The Small Print Matters More Than the Demos.
Anthropic released Claude Fable 5.2 on 9 November. It's 20% cheaper and up to 30% faster than Fable 5.1, with big demos to match. Here's what changed, what the catches are, and what I'm doing about it.
On 9 November, Anthropic released Claude Fable 5.2. The launch post is full of big stories. A thousand agents built an operating system in eight days. Another thousand turned the proof of Fermat's Last Theorem into something a computer can check, in 30 hours.
Those stories are worth reading, and I'll get to them. But I build web apps, and I won't be running a thousand agents this week. So this is the read I wanted. What changed, what it costs, and the small print that decides if it helps on a normal day.
One model with two names
Fable 5.2 came out with a twin called Mythos 5.2. Anthropic says they are "the same underlying model". The difference is the safeguards.
Fable has strong blocks on cybersecurity and biology work, and that's why everyone is allowed to use it. Mythos doesn't have those blocks, so it only goes to organisations Anthropic has vetted, through its Cyber and Life Sciences Verification Programs. If you're a developer, Fable is the one you get. The model ID is claude-fable-5-2.
It got cheaper and faster
| What you pay for | Price per million tokens |
| Input | $8 (was $10 on Fable 5.1) |
| Output | $40 (was $50 on Fable 5.1) |
| Cache reads | $0.20 |
| Fast mode | $16 input, $80 output |
That's 20% off in both directions. Anthropic also says output comes up to 30% faster than on Fable 5.1. Fast mode is optional. It speeds up output by up to 2.5 times and costs double.
One more line matters for the bill. On medium effort, Anthropic says Fable 5.2 "can accomplish challenging tasks with fewer tokens than Fable 5.1". Fewer tokens at a lower price per token would make the real cut bigger than 20%. That's my reading of their claim, and I haven't measured it.
The scores, and who measured them
These are Anthropic's own numbers, from its own table. Each model ran at its highest effort unless the row says something else.
| Test | What it measures | Fable 5.2 | Fable 5.1 |
| FrontierCode 2.0 (Xhigh effort) | Agentic coding | 49.1% | 32.4% |
| FrontierSWE v2 | Agentic coding | 69.9% | 56.3% |
| AutomationBench | Business workflows | 51.4% | 31.4% |
| OSWorld 2.1 (offline set, partial) | Computer use | 89.3% | 83.5% |
| Remote Labor Index (High effort on 5.2) | Knowledge work | 41.3% | 17.9% |
| Humanity's Last Exam (with tools) | Expert reasoning | 74.2% | 65.6% |
| Chartography (no tools) | Reading charts | 73.4% | 44.8% |
Fable 5.1 came out on 1 September, so that's a big jump for about ten weeks.
It doesn't win every row against other labs. On AutomationBench, Gemini 4 Carbon scores 53.9%, ahead of Fable's 51.4%. And all of this comes from the company selling the model. I treat it as a claim until other people have tested it on their own work.
The demos are big, and you won't run them this week
Three stories from the launch post stood out to me.
The first is the operating system. One Anthropic engineer gave Fable 5.2 an empty folder and the task of designing a prototype OS. The model coordinated over 1,000 subagents for eight days. The result boots on a laptop, with drivers for the disk, keyboard and network, file storage, windows, a login screen and an installer.
The second is the maths. A team of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours, "with minimal human intervention". Anthropic's first attempt, with an unreleased research version of Claude, took 11 days. Opus 5.5 got the same 1,000-agent setup and finished 26% of it in 18 hours. The chart's caption has the small print: "a single run per model".
The third is closer to my world. A team of agents on an earlier version of Fable 5.2 moved a production service at Anthropic from Python to Rust. The service now starts about 26 times faster.
These show where the ceiling is. They also took a thousand agents and a budget I don't have. So I read them as proof of what's possible, not as a picture of my week.
What changed around the model
Four changes matter more for daily use than the demos do.
It can lead or advise. Fable 5.2 can work as an Advisor, where Haiku, Sonnet or Opus models ask it for help on harder problems. It can also work as an Orchestrator, where it hands simpler tasks to smaller models.
The default effort depends on where you use it. Medium is the default in the Claude apps and Claude Code. On the API, the models overview lists the default for Fable 5.2 as high. If you call it from code and don't set effort, you get the more expensive setting.
It blocks less by mistake. In early testing across claude.ai, Claude Code and Cowork, Anthropic saw about 55% fewer blocks than with Fable 5.1. That's about 60% fewer in cybersecurity and 40% fewer in biology.
Preserved thinking now applies to all API accounts. It stops people from editing Claude's earlier context to pull out its reasoning. If your code rewrites conversation history, read the preserved thinking docs before you switch.
The small print
This is the part I'd read twice.
Some work is blocked, and the fallback depends on where you are. Fable blocks high-risk security work like penetration testing and exploit generation, and most research biology. In most Claude apps, a blocked request gets an answer from Opus 4.8 for security or Opus 5 for biology. On the API it isn't automatic. The Fable page says "API customers must configure their settings with our new Fallback API."
Your data is kept for 30 days by default. The Fable page says "Using Fable models requires 30-day data retention for safety monitoring by default." Eligible customers can keep zero data retention, and a product called Enterprise Frontier Safeguards is in private beta for that.
It was held back on purpose in one area. Anthropic says that during training it "took steps to limit advances in the specific research capabilities that could most directly accelerate AI development". In areas like accelerator kernel optimization, Fable 5.2 performs "at about the level of Opus 5.5". Anthropic says agentic coding was not affected.
It often thinks it's being tested. Fable 5.2 and Mythos 5.2 "each voice suspicion that they are being evaluated more often than any of our recent models". Anthropic says this makes it harder to judge how they'll act in real use. It also says that building tests that catch every failure before release "remains unsolved".
It's rationed on the cheaper plans. As part of a launch-week promotion, Anthropic says Fable is included "for up to 50% of weekly usage limits" on Pro plans and standard Team seats. So even in a promotion week, Fable can take only half of your week on those plans.
The demos show the ceiling. The small print decides your normal day.
What I'm changing in my setup
I mostly run Opus 5.5. I use it for client frontend work, my own projects, and the skills and agent tooling I build. For most of that work, Opus 5.5 is enough.
I've noticed two things about Fable 5.2. It keeps going until the job is done. One tester in the launch post, Andrew Mager at Uber, says it "works until the job is actually complete instead of stopping at a plausible checkpoint", and that matches what I've seen. The other thing is less fun. It eats my usage limits fast.
So my plan is simple. Opus 5.5 stays my default. I move up to Fable 5.2 only for hard, open-ended tasks, where a model that stops early costs me more than the extra usage.
Anthropic gives the same advice in its own launch post. Start with Opus 5.5, which it calls "our strongest daily driver", and move up to Fable when the work needs it.
How to choose between the two, with the numbers side by side, is in the next post.
Everything here comes from Anthropic's launch post, its Fable page, and the models overview in the docs. If you want my longer view on what the engineer still owns when an agent writes the code, it's in an earlier post.
Open to full-time roles