Anthropic has finally released the safe version of the Mythos model, Fable-5.

Many details about the parameters have already been discussed online, so I won't repeat them.
What’s more worth noting is the real task testing done by a third-party programming tool, Augment Code.
This set of tests ran a total of 489 programming tasks, and the results are quite interesting:
Fable-5 clearly leads in overall performance and accuracy.
Overall score +0.224, accuracy +0.191, definitely the top tier so far.
But another detail is also very important: GPT-5.5 still comfortably outperforms Opus-4.8.
Overall score GPT-5.5 is +0.164, Opus-4.8 is +0.128;
Accuracy GPT-5.5 is +0.141, Opus-4.8 is +0.092.
This also explains my recent feeling: after Opus-4.8 came out, I didn't feel it was significantly stronger than GPT-5.5—at least in practical programming tasks, this feeling is not an illusion.
A more realistic concern is cost.
Although Fable-5 is powerful, its token consumption and costs are also high: about 14.6k tokens per task, with a cost of $3.09 per task;
by comparison, GPT-5.5 uses 7.5k tokens and costs $1.52.
The power is real, but so is the expense.
So in the end, I still say: looking forward to GPT-5.6 arriving sooner.
If Fable-5 can only be used for 10 days in the subscription plan, and afterward must be called at the API’s regular price,
then it’s likely not a daily productivity tool for ordinary users, but rather a “luxury model” for a few people and scenarios.
The use of AI models may really start to be tiered.