Models, Research & Prompt ·
Claude Fable 5 Is Here
English translation of the Chinese original. This version is generated for international readers and may be refined over time.
English translation of the Chinese original. This version is generated for international readers and may be refined over time.
Date: 2026-06-10
Yesterday, Claude Code might have released news of a new model, but it's coming.
I'll give you the background. When Opus 4.8 was last released, the official forecast was to bring Mythos to all clients within weeks. Mythos Preview has not been officially released and has only been commercialized for a small number of security agencies and infrastructure plants on the grounds that it is too capable of network security to be misused.
This time, Claude Code's approach was to break the same bottom model into two names. Mythos 5 continues to remain in the hands of the original group of partners and is not reached by ordinary users;
Fable 5 added a safety cataloguer and released it to everyone.
The name also has to be well known, can come from Latin fabula, and mythos is the same language as Greek. One translates a fable, one translates a myth, and this is for everyone, called the fable.
Take a look at the assessment data.
The header itself has the amount of information: Claude Mythos 5 and Fable 5. The small word says that the difference between the two models is between one and three points, and the tall one is on the table.
Compared to Opus 4.8, full-line wins, a few of them pull out. SWE-Bench Pro 80.3% for 69.2%. Frontier Code 29.3% for 13.4%, more than doubled. Space reasoning Blueprint-Bench 38.6 per cent for 14.5 per cent, nearly three times.
Compared to GPT-55, most of it's on the prowl. SWE-Bench Pro 80.3% against 58.6% and 21 points.
The FrontierCode is more visual. Fable 5 isn't just a pile of tokens, but it's pretty on the cost curve.
But Terminal-Bench this much more I say. The last time Opus 4.8 released me, it couldn't do GPT-5.5. This time it says 88.0%, which puts 83.4% of Codex CLI under pressure, but I have to tell you, it's with an asterisk.
What's an asterisk?
Underword: Project with an asterisk because Fable 5 has a security downgrade on issues related to cybersecurity and biology and actually scored closer to Opus 4.8. So 88.0% is the score of Mythos 5, and you've got this Fable 5 job, probably 82.7% up and down, even with Codex CLI.
The most typical is Expluit Bench, 78.0 per cent of the watch, and Opus 4.8, 40.0 per cent, which looks like hanging. But it's also with an asterisk, and you really take Fable 5 to run, and the scores fall back around 40%.
More interesting than Mythos Preview. SWE-Bench Pro 80.3% for 77.8%, OSWorld 85.0% for 85.4%, HLE tools 64.5% for 64.7% - the latter two Previews are far more.
To be honest, Fable 5 and Preview, the one thing that scared the cyber security unit so badly, was a piece of something that won and won.
But what needs to be noted is that the de-escalation mechanism has also said: ** Ask sensitive questions and answer your questions quietly to Opus 4.8. The limitations are mainly reflected in three categories - cybersecurity, biology and chemistry, and model distillation. **
The official also gave an early data: more than 95% of Fable sessions did not fallback. In other words, if you're doing normal code writing, changing files, reading newspapers, running long missions, most of the time you get mythos.
Anthropic put the escape on the stage this time. In the internal evaluation, the auto-red team will try again in 400 rounds, roll back and continue rounding; the outside bug bounty exceeds 1,000 hours and does not have access to United jailbreak. Officially, 30 open jailbreak technologies were measured by an external partner and Fable 5 was a zero-cooperative request for harmful single-wheel network security.
The official statements cannot be fully believed, after all, are official self-censorship and partner testing.
But the direction is set: Fable 5 will not refuse to carry out your mission. Instead, this task automatically downgrades the model to Opus 4.8.
In addition to cyber-security, the official space devoted to biology and chemistry was considerable.
Mythos 5 has made an AAV-related assessment. In short, the experimental properties of predicting the virus's shell. Officially, these candidates came from Dyno Therapeutics, who did not have specialized training models to do the job, but the Mythos-class model exceeded the specialized protein model model baseline by biological reasoning.
There are two sides to this.
On the one hand, it's a good thing that models are really starting to work in the direction of gene therapy, drug development. Officially, the internal protein design specialists used Mythos 5 to speed up some of the drug design processes approximately 10 times; of the 14 protein target points, 9 produced strong candidates who are continuing research.
The other side is risk: the same ability to do bad things in the wrong hands.
So Fable 5 will temporarily fall back to Opus 4.8 on most biological and chemical requests. Officially, there will be a biodirected access program: remove biological and chemical restrictions, but retain network security restrictions.
And that's what Anthropic is thinking right now, tearing down the power, and also the access.
The price is $10/$50 per million token. Gives two references: twice the standard price of Opus 4.8 ($5/ $25), which is exactly the price of Opus 4.8 fast Mode; less than half of the $25/$125 preview of Mythos Preview. Context 1M, Max output 128K.
There's a time to focus on the subscription here. Claude Code has a line of small words on his model selection: Incruded until June 22. Before June 22, Fable 5 was included free of charge in Pro, Max, Team and Enterprise subscriptions; the 23rd was removed from the subscription and can only go API on a volume basis. Officially, when sufficient capacity is available, the subscription is returned.
** Which means the next 12 days is a window period, and I'm on the list. This 12-day limit, straight to full Fable 5 prison. **
Speaking out of context, I saw Han in an official video from Claude. Isn't that in some way A might be better for China?
That's it for today. I'll see you next time.
- Officially issued thread:https://x.com/claudeai/status/2064394146916229443
- Models and pricing documents:https://platform.claude.com/docs/en/about-claude/models/overview