The short answer
● Same model, same price — Both are $8/M input and $40/M output. The weights are the same.
● Fable 5.2 is the one you can use — Open to every user. Safeguards block most cybersecurity and biology tasks.
● Mythos 5.2 needs an invitation — Fewer restrictions, but only for vetted cybersecurity or biomedical organisations.
Anthropic shipped two models on Monday that are, by its own description, built on the same underlying model. They cost the same. The difference is entirely about what each one will agree to do, and who is allowed to ask.
|
Fable 5.2 |
Mythos 5.2 |
| Input price |
$8 / M tokens |
$8 / M tokens |
| Output price |
$40 / M tokens |
$40 / M tokens |
| Batch API |
$4 / $20 |
$4 / $20 |
| Who can access it |
Every user |
Vetted organisations only |
| Cybersecurity tasks |
Mostly blocked |
Permitted |
| Biology tasks |
Mostly blocked |
Permitted |
| Speed vs 5.1 |
Up to 30% faster |
Up to 30% faster |
| Previous version price |
$10 / $50 |
$10 / $50 |
Why two models instead of one
A single model with safeguards strong enough for public release is too restricted for the people who need it least restricted — security researchers testing exploits, biomedical labs working with pathogen data. A single model loose enough for them is not something you hand to everyone. Anthropic resolved it by shipping both and gating the second one on who you are rather than what you pay.
Which one applies to you
- You are building a product on the API → Fable 5.2. Mythos access is not something you apply for casually, and for ordinary work the safeguards never fire.
- You work in security research at a recognised organisation → Mythos 5.2 is worth the vetting process. Fable will refuse a meaningful share of what you ask.
- You are on Fable 5.1 today → move. Same capability class, 20% cheaper, faster on coding.
- You are cost-sensitive and not doing frontier work → neither. Claude Haiku 5.5 is $0.10 against Fable’s $8, an eightyfold difference.
The honest limitation
Anthropic has published no benchmark scores for either model. The claims available are 20% cheaper, which the pricing page confirms, and up to 30% faster, which it does not. Treat the speed figure as a vendor claim until independent testing lands.