Releases

Claude Fable

Anthropic has released their highly anticipated Claude Mythos model under the name Claude Fable. It is not the exact Mythos Preview model that people got access to as a part of Project Glasswing, but rather a model from that family (different finetune of the model).

The main difference from Mythos to Fable is the increase in safety guardrails to prevent the model from engaging in harmful activities like biology and cybersecurity.

To be rather blunt: this is the best LLM out there right now for text based tasks, if you are willing to foot the rather large bill for it.

It is the best at frontend design, coding, and agentic tasks. Anthropic clearly does not care about vision capabilities, as Fable is the about as good as the mid size Qwen 3.6 models on vision benchmarks.

As for pricing, its 2x the price of Opus. It also uses more tokens than GPT 5.5 xhigh on similar questions, so expect it to be 3-5x more expensive in the real world than GPT 5.5. OpenAI continues to be far ahead of every other lab at squeezing the most intelligence out of the fewest tokens.

Model	$ per million (input)	$ per million (output)	Tokens per second
Claude Fable 5	$10	$50	40
Claude Opus 4.8	$5	$25	60
GPT 5.5	$5	$30	51

There is also the issue of the safety filters. If you are a biologist, sorry to say but you won’t be able to use Fable whatsoever. Anthropic has decided the best way to prevent any bio safety risks is to make the model refuse anything bio related, including “Is mitochondria the powerhouse of the cell?”. They also have safety filters for cybersecurity tasks (Mythos originally got famous for being very good at hacking things) and also advanced machine learning topics.

The machine learning filter gathered the ire of the community specifically. For all other topics, when a safety filter is triggered, the user would be informed that they had triggered the filter and were being rerouted to Opus. For machine learning tasks, Anthropic thought they could sneak in the fact that the user would not be notified, thereby sandbagging the results and silently giving the user worse performance, which was seen as very underhanded and actively hostile towards the community.

All of this is a moot point right now, since 3 days after its release, Fable access has been removed for everyone, following an export control directive from the US government that suspended all access to any foreign nationals (anyone not a US citizen) both inside the US and out. This includes any Anthropic employees that are foreign nationals as well. The only way that Anthropic was able to abide by this ruling was to turn off access for all customers.

What appears to have happened is that one of the enterprises using Fable (supposedly Amazon, a major Anthropic investor) was able to jailbreak the model and bypass the safety filters for cybersecurity, and reported this to the US government. The US government, upon hearing this, and the fear mongering that Anthropic propagated with the initial Mythos announcement about how dangerous of a hacking tool it could be, decided that they did want any foreign adversaries to have access to the model, and issued their decree.

I don’t think Anthropic is necessarily upset about this. They have always wanted nobody else other than them to be working on powerful AI systems, and this new ruling sets the precedent that at a certain capability level you should expect the US government to step in and make things messy. This could effectively limit the companies who can work on AGI in the US to only the big players that have the money and lobbying power to convince the US to allow them to continue working on these systems. This would mean that only the likes of Anthropic, OpenAI, and Google may be able to effectively work on these systems in the future.

This has caused quite a bit of chaos in the AI world. OpenAI has seen a large shift as users head to their platform in response to Anthropic’s treatment of the ML community. Many individuals and companies have also announced that they will begin working on their own models and finetunes, to prevent losing access to models in the future, and also not have a third party dictate what they can or cannot use a model for.

I expect Anthropic and the US to sort out their issues and Fable to return in the following weeks. This is clearly an inflection point in the AI race, and we will be feeling the effects of this for a long time to come.

Quick Hits

Kimi K2.7 Code

Speaking of open source models, let’t talk about MoonshotAI’s new Kimi K2.7 Code model. This is a coding specific finetune built on top of their K2.6 model.

I think I was a bit harsh in my Kimi K2.6. As time has gone on it seems to be a decent model, roughly in line with all of the other frontier Chinese models.

They also seemed to have cleaned up their benchmaxxing a bit, as I think pretty much everyone had caught on that the scores they were showing were meaningless and were not driving adoption. Now the scores that they show are clearly below the likes of GPT and Claude, but also a nice bump up from the previous version.

They also have worked on getting token usage down, reducing token use by 30%. This model has flown under the radar a bit due to all of the Fable drama, but from the bits that I have seen from it I would say it’s among the best open source coding models out there right now.

Finish

I hope you enjoyed the news this week. If you want to get the news every week, be sure to join our mailing list below.

Shangchen Bathhouse in Hangzhou, China, designed by Wei Zhixue.