Anthropic released the strongest AI model it has ever opened to everyone. In the same announcement, it told most people to start with a different one. That detail says more about how to use AI at work in 2026 than any benchmark chart on the page.
Anthropic is the company that makes Claude, the AI assistant many developers use for coding, writing and research. On November 9, 2026, it announced Claude Fable 5.2 and Claude Mythos 5.2. The headline number is a record. The part that will change your bill is quieter: the strongest model, set to think less, matches Opus 5.5, Anthropic’s strongest everyday model, for a fraction of the cost.
So the useful question is “for which tasks, and at what setting?”. At the end of this post I give you my answer: which model I would use for each kind of task, and at which effort level.
What Anthropic launched on November 9
Two names, one launch:
- Claude Fable 5.2 is for everyone. It is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure, and on the Claude Platform with the model id
claude-fable-5-2. - Claude Mythos 5.2 is for vetted organizations only, through verification programs. Anthropic calls it its most capable model for cybersecurity engineering.
Anthropic positions both for “the most ambitious, open-ended technical work”. And then, in the same post, comes the line that caught my attention: “For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver.”
A company telling you not to start with its newest, most powerful product is rare. It is also good advice, and I will come back to why.
One model, two names: Fable and Mythos
This is the part that confuses people. Anthropic says it directly: “Claude Fable and Claude Mythos are the same underlying model.”
The difference is the safeguards. Safeguards are the filters that block dangerous requests, for example help with building a cyberattack or a biological weapon. Fable 5.2 ships with strong cybersecurity and biology safeguards, and that is the version Anthropic makes available to all users. Mythos 5.2 runs without some of those safeguards, so it is restricted to organizations that apply and get approved for beneficial work in cybersecurity or biomedical research.
For almost every reader of this blog, Fable 5.2 is the one that matters. If you are not in a security or biomedical lab with a verified program, Mythos is not on your menu. It is the same model with fewer filters, so you lose much less than the name suggests.
What the strongest model can do, according to Anthropic
Every number in this section comes from Anthropic’s own announcement. Several are on new benchmarks (a benchmark is a standard test used to compare AI models). Anthropic also notes that the external organizations it worked with did not review the contents of the post. Read them as the vendor’s claims, which they are.
With that said, the claims are strong:
- Benchmark record. On FrontierCode 2.0, Fable 5.2 at its Xhigh effort level reaches 49.1%, which Anthropic calls “the highest score of any released model we’ve tested”.
- Fermat’s Last Theorem, formalized in 30 hours. Anthropic ran 1,000 Fable 5.2 agents (an agent is an AI that executes steps on its own) and they fully formalized the theorem in 30 hours. Formalizing means rewriting an existing proof so a computer can check every step. They did not discover a new proof. The first full formalization, done by an unreleased research version of Claude, took 11 days. Opus 5.5, with the same 1,000-agent setup, completed only 26% in 18 hours.
- Hardware. Fable 5.2 designed three working electronic devices: a voice hub, a USB speakerphone and a handheld virtual pet.
- A prototype operating system. Coordinating over 1,000 subagents for eight days, around the clock, it built a prototype OS that boots on a laptop.
And one number about Claude in general, not Fable 5.2 specifically: “Over 90% of the changes merged into Anthropic’s core codebase are now written by Claude.”
Notice what these examples have in common. They are long, open-ended jobs with many steps and many agents. Nobody needed Fable 5.2 to rename a variable.
Price and speed
Prices for AI models are quoted per million tokens. Tokens are the small pieces of text the model reads (input) and writes (output); a token is roughly part of a word.
| Input (per million tokens) | Output (per million tokens) | Speed | |
|---|---|---|---|
| Fable 5.2 | $8 | $40 | up to 30% faster than Fable 5.1 |
| Fable 5.2 fast mode | $16 | $80 | token generation up to 2.5x faster |
Anthropic describes the standard price as lower than before, but the page does not give the old number, so I will not guess a percentage.
Fast mode is simple math: double the price for up to 2.5 times faster output. That is a good deal in a few specific situations and a waste in most others.
For people on subscriptions, Anthropic says Pro and standard Team seats get Fable included “for up to 50% of weekly usage limits”, plus “one to three Fable resets based on how long they’ve subscribed”. The details are on a separate page, so check your own plan before you count on it.
The twist: the effort dial matters more than the model name
Here is the part that is easy to miss.
Fable 5.2 has an effort level: how long the model “thinks” before it answers. The levels are Low, Medium, High, Xhigh and Max. Medium is the default in the Claude apps and in Claude Code (Anthropic’s coding tool for the terminal). Most people never change it.
Now look at what Anthropic reports when you turn that dial down:
| Benchmark | Fable 5.2 effort | Result vs. Opus 5.5 | Cost per task |
|---|---|---|---|
| FrontierCode 2.0 | High | 45.3% vs. Opus 5.5’s best 45.5% | roughly half |
| FrontierCode 2.0 | Xhigh | 49.1%, the record | not stated |
| Humanity’s Last Exam | Low | matches Opus 5.5’s top score | less than one-tenth |
| WANDR | Medium | matches Opus 5.5’s top score | about 40% |
Read the first line again. One step below the record, Fable 5.2 lands about where Opus 5.5 lands at its best, for roughly half the cost per task. On Humanity’s Last Exam, a set of very hard expert-level questions, the lowest effort setting already matches Opus 5.5’s top score.
So the record headline (49.1%) is real, but it lives at Xhigh. The practical news lives at Low, Medium and High.
This is why I think Anthropic’s advice to start with Opus 5.5 makes sense. The newest model should not be your default. The skill that pays now is choosing the model and the effort for the task in front of you, and moving the dial only when the result falls short.
Fewer refusals for legitimate work
If you do security work with Claude, you may have hit a block on a perfectly normal request. Anthropic says this improved: in early testing across claude.ai, Claude Code and Cowork, it saw about 55% fewer safeguard blocks with Fable 5.2 than with Fable 5.1. By area, that is about 60% fewer in cybersecurity and about 40% fewer in biology.
When a request is still blocked, Anthropic says it is answered by an older model instead: Opus 4.8 for cybersecurity and Opus 5 for biology. Worth knowing if an answer suddenly feels weaker than you expected.
On safety in general, Anthropic assesses the overall risk of unsafe behavior as “low (as with our other recent models)”. One detail from the announcement I found interesting: the new models “voice suspicion that they are being evaluated” more often than recent models. That matters for anyone reading benchmark results, including the ones above.
My verdict: which model and which effort for each task
This is my opinion, based on Anthropic’s numbers and on how I judge any tool at work: by cost and delivery.
| Kind of task | Model I would use | Effort to start | When I would go up |
|---|---|---|---|
| Daily coding: a bug fix, a small feature, a code review | Opus 5.5 | its default | Rarely needed; switch to Fable 5.2 only if it keeps failing |
| Writing, summaries, emails, documentation | Opus 5.5 | its default | Almost never |
| Long, open-ended technical job: a migration, a big refactor, research across many files | Fable 5.2 | Medium | High when the result falls short |
| The hardest problems, where the best possible answer is worth the cost | Fable 5.2 | High | Xhigh or Max, deliberately, with a budget in mind |
| Anything urgent where waiting costs more than tokens, like a live production incident | Fable 5.2 fast mode | Medium | Only for the duration of the emergency |
The rule behind the table is short: start with the daily driver, move to the strongest model only for long and open-ended work, and climb the effort dial one step at a time. Most of the value Anthropic reports for Fable 5.2 shows up below the maximum setting, so starting at the top means paying for thinking you did not need.
Fast mode deserves its own warning. It doubles the price. In my view, it only makes sense when a person or a client is waiting on the answer and every minute has a real cost.
Before you switch anything, run through this:
- Is this task long and open-ended, or is it a normal daily task? If it is normal, stay on Opus 5.5.
- Did you try the default effort first? Go up one level only after a result that falls short.
- Do you know what a task costs you today? Without that number, “half the cost” means nothing to your budget.
- Are you comparing your own results, or only Anthropic’s benchmarks? Test on two or three of your real tasks before deciding.
The benchmark record will be beaten by the next launch. The habit of picking the model and the effort for each task is what keeps paying off.
And you: do you pick a model per task, or do you still run everything on the biggest one?
Sources
- Anthropic, “Claude Fable 5.2 and Claude Mythos 5.2”, November 9, 2026: anthropic.com/claude-fable-and-mythos-5-2. All benchmark figures, prices, speeds, safeguard numbers and examples in this post come from this page and are Anthropic’s own reported results.
- Anthropic, Claude Fable 5.2 and Mythos 5.2 system card: anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card (linked for further reading; not used for any figure here).
Leave a Reply