Claude Fable 5.1 and Claude Mythos 5.1

Interesting that Fable produced the Venus elevation map by training a neural net to generate it. I wonder what the prompting looked like to make that happen, I.e. was this a spontaneous discovery or the result of a specific request.

Yeah but haiku 5 when?

There's now a 40X discount in the cache input pricing instead of 10X.

This seems to point to them having achieved some kind of optimization in attention mechanism perhaps along the lines of DeepSeek V4, which had a similarly high discount between cache input and normal input.

In real world use, the savings should be quite noticeable. For example, you can now use the model at 800K tokens context window at the same cost efficiency as the previous model at 200K tokens context window.

Interestingly, Claude’s output is now actually readable with Fable 5.1. Pretty sick.

Did we get thought traces back? If no, it's useless.

Worth noting: Claude Fable 5.1 and Mythos 5.1 are Anthropic’s first models to watermark text outputs.

> In part, this is because Fable 5.1 can now be used to discover software vulnerabilities—though not to develop exploits for them

Generally once an exploit chain is described, developing the exploit is trivial.

If you're so inclined, discover the exploits using Fable 5.1 and then give that exploit to a model that doesn't have such compunctions (e.g. local LLM or an uncensored cloud model / model that's easier to jailbreak). I don't think Anthropic is really mitigating here anything in the real world other than PR narratives where media can report "Anthropic's model was used to develop the latest cyber attack".

I find myself wondering how much of the writing style is based on financial incentives?

When paying by the token, don't the labs have a strong incentive to make the model as verbose as possible?

ah, the smell of new model.

OpenCode+GLM-5.3 (and 5.2) for three weeks solid. Im so happy I made it out

This time it came with a usage reset

> Data retention. Our new system of Enterprise Frontier Safeguards (EFS) gives customers complete privacy (the same as a zero data retention policy) while still being state-of-the-art at preventing adversarial use. EFS works by storing data in cloud infrastructure controlled entirely by the customer, not Anthropic. It will be made available to enterprise customers in phases, beginning later this fall. Until EFS is available, eligible customers will be able to use Fable 5.1 with zero data retention.

This is interesting. I wonder if customers will be allowed to create an auto expiry for their own data to prevent future subpoenas. That’d be a treasure trove for discovery.

Unless these people start offering free, unlimited inference for a cautionary period so we can test the new model without an up-front (re-)investment, I am not touching this load-bearing pile of neuralese spew with a ten thousand token pole.-

I'm having a very hard time finding mention of token-generation speed.

Even with discounted cache, their prices remain way above everyone else but not necessarily the results.

What exactly is the premium that you're getting for paying these prices?

Why aren’t these models available on subscription plans?

I tried the old fable and it didn’t seem worth paying for. It still made errors like Opus does so I might as well use the included model…

I don't know how I feel when all the documentations are written by AI for humans.

AI to AI doc share: sure, do what you please.

AI to human: please make it legible and flowly.

example, "Every thinking block records which model produced it, and it's preserved in one direction only: Claude Fable 5.1 reads earlier models' thinking blocks, and no earlier model reads Claude Fable 5.1's." is a very Claude-isk way of writing. Choppy, long, and lacking flow.

Is Fable 5.1 still actively downthrottling the reasoning when questions relate to frontier ML questions, like it did with 5.0?

This coupled with verification primitives will be quite compelling. we really have to start reimagining existing systems and processes from the ground up.

Great,excited to use these models

Will it still refuse my mitochondria questions?

Unfortunately at the moment the model is very quickly burning through plans, single session with 3-4 subagents, none using Max or Xhigh, mostly medium + some High can burn through 5h limit within 20-40 minutes of usage.

"with cache reads at a quarter of the cost"

OK, I think that's what they meant when they suggested reduced extra promo usage will not sting this much.

> Forced tool use is not supported

That seems unfortunate for 3rd party integrations that expect stable output - what that really necessary ?

Anthropic, the company employing "treat them mean, keep them keen" as a marketing tactic. Pass.

Going to hold off a few days until I adopt it, lets see what the general consensus develops as. Regretted jumping over day one for 5.0. The caching thing seems the most useful, but doesn't change anything for my subscription.

I use Claude Design heavily, I wish these charts show "10% better at picking a color" or laying out an app. Maybe it's hard to build a good visual design test. Claude's good at layouts but not the colors or smaller design details.

My only concern is that sooner or later the best models will be priced out of my ability to pay.

I have been happy with Fable 5, it has done great work for me so far. Very excited to try out Fable 5.1 and see what differences and improvements there are.

Unfortunately isn't included in subscriptions and requires usage credits...

Looks like the API is nerfed to mitigate some recent thinking extraction attacks.

I wonder to what extent this will make the automatic Fable-to-Opus downgrade give worse results.

I'm afraid watermarking could restrict applications where LLMs can be safely used to assist with writing. If I write something myself and use an LLM to proofread it, without watermarking I can confidently say that corrections done by LLMs are small and insignificant enough to claim that the text is still authored by me, not by the model. With watermarking, however, I will never be sure if the result will not be flagged as AI generated, even if the AI contribution is very minor.

I have to say, I am quite frustrated with Anthropic lately. I so badly want to use Fable to work on a side project of mine, which I used to do previously with no issues. But lately, they must have made some classifier change because it keeps hitting their stupid, overly-hyper-aggressive safeguard due to 'general_harms'.

Guys, listen to your feedback please. I hadn't used OpenAI products in quite a while until this issue came around. They seem to have MUCH smarter safeguards than Anthropic does.

I'm confused about Anthropic's pricing. Can anyone explain why Sonner 5 is $2/MTok in and Sonnet 4.6 is still $3?

"Content provenance" seems to be activated with this model.

I noticed they reset the usage and I was kind of happy because this week it was using my quota much faster; I assumed they fixed that. Apparently it is for the celebration of 5.1?

Are we getting a new Opus 5.1, then?

Am I alone in not prioritizing the quality of prose produced by my coding agent? My foremost and almost only concern is how well it can engineer software.

The average company and definitely average Joe will never be able to afford is ludicrously expensive model. Do not use this in a corporate/startup environment unless you have endless VC cash.

Cool. I’ve realized though that I don’t really need better models anymore. SOTA is good, I just want them faster/cheaper now.

Fable 5.1 seems to be the first model who can accurately draw an airbus a320 in 3D space given a set of limited tools (a brush with params color, size hardness and xyz coords): https://youtube.com/shorts/vyHsMqop2yw

I yearn for a model that can churn through claude text and write sensible text. so far gemini is pretty good at that, even in the low variant

If Anthropic thinks Opus 5 is very good, it is a window into how insular their culture is. I find it far, far behind Sol. It’s downright annoying to use.

Almost finished my weekly limit today! I am more excited from the usage reset!

Not until you stop being cheap and let pro users use fable under their existing paid subscriptions.

The safeguards and required extra retention is still not gone. Further more they are working to create separate tiers of access with the new biology program instead of giving everyone equal access to AI. Anthropic once again are showing they can not be trusted.

The breaking API changes are frustrating, especially the one that removes forced tool use.

"Price. Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads, wherever usage is billed by token. This is because we’re reducing our pricing on cache reads (where the model reads inputs that have already been processed and stored). For highly agentic work, the savings will often be much larger—up to approximately 45%."

They show this off, but artificial analysis contradicts the statement. Fable 5 cost $3.14 per task, while 5.1 cost $3.69 -- around a 15% jump in pricing.

https://artificialanalysis.ai/

These, IMO, are marginal improvements for a more expensive model. I stopped using Claude ~3 months back; its outputs are too jargoned, it makes architectural decisions that are not right, and it's incredibly pricey for what it is. Each decision it makes, it acts as if a problem as major as world hunger has been solved. And the overly verbose code comments, strange commit descriptions, duplicate code, and slop it generates -- which I know is not specific to Fable -- is just too much for me.

I found the best is to use something like Deepseek V4 Flash -- with a fast TPS provider -- and work on the code myself. For agentic work with computer use, GLM 5.3 flash with Hermes Desktop works well.

The most remarkable thing here is just how close Opus 5 is on most of these benchmarks.