Mythos 5 and Fable 5: My Honest Take
Anthropic just shipped a brand-new family of models that sits above Opus: Mythos 5 and Fable 5. They were released less than twenty-four hours before I recorded this, so this is not a deep benchmark study; it is my honest first take on where these models actually make sense, where the pricing does not, and why most businesses should keep their budget on Opus and Sonnet.
1. What changed in Anthropic's model family (0:10)
Mythos 5 and Fable 5 are a new architecture, not another Opus revision. Mythos has existed since Project Glasswing as a private model for a few companies and users. What is being released broadly now is Fable 5, which uses the same class as Mythos but ships with a few more restrictions, making it safer for general use. Mythos 5 preview is also available, but only to a small set of companies.
2. Pricing and the business-use reality check (0:54)
The headline you need: this pricing is roughly a two x increase compared to Opus 4.8 and its predecessors, with output running at fifty dollars for every single million of tokens. These models consume monstrous amounts of tokens, so at scale this gets extremely expensive. For business use cases, I do not see it: Opus 4.8 and Sonnet already get most business things done just right.
3. Benchmarks and the cybersecurity jump (1:38)
Always take first-party benchmarks with a grain of salt; they come from the same company releasing the model. Across the board you see increases, but compared to the models that came right before, the gains are not dramatic, except in one very precise section: cybersecurity. For Fable 5, that is nearly a two x jump over Opus 4.8, and that is where I think the true value of this model class lives. It skews toward heavy developers, back-end developers, and cybersecurity work, not business users.
4. Longer agentic work and cost per task (2:50)
Further down the charts there is a real increase in the independence of these models: they can work for longer stretches on a task without you. That again points at developers. The flip side is that the average cost per task you hand them also goes up, so factor that in before moving serious workloads over.
5. Vision, context quality, and the demos (3:24)
Vision has been one of the limitations of AI generally, and Gemini has probably been the best at it. So the vision improvements here are genuinely interesting for design work, image analysis, and video analysis. Anthropic even shows Fable 5 beating Pokemon Fire Red using only computer vision, which is fantastic.
On context: they did not increase the size of the context window, and that does not really matter. What matters is the quality of that window, meaning how well the model remembers what you told it within a single conversation, and that improved.
One demo I would file under marketing: the solar system simulation that quote-unquote predicted a solar eclipse. We can already predict solar eclipses with extreme accuracy millions of years out, because we have known those equations for a very long time.
6. Who these models are really for (4:54)
Researchers, cybersecurity developers, and hardcore power users of Claude. Anthropic cites research use in biology and chemistry. I have not yet met a researcher actively using this model class, but it is clearly the audience they are pushing toward. They also increased what they call alignment: the model's predisposition to give you good, human answers grounded in morals and ethics.
7. Mythos vs Fable: the safeguards (5:39)
The main difference between the two is safeguards. Fable 5 has more of them; Mythos 5 can behave in a quote-unquote more dangerous manner. Beyond that, they are oriented toward the same pricing and the same level of usage.
8. Data retention and compliance limits (6:33)
Here is genuinely bad news: Mythos-class models require 30-day data retention for all traffic. If you need to comply with standards such as ISO 27001, which requires zero-day data retention, you cannot use Fable 5 or Mythos 5 yet. Power users and developers who need that level of trust should sit this one out for now.
9. The access window and the token burn warning (7:27)
Right now, every regular user gets access to Fable 5. That changes after a limited window of days, when it moves to usage-based pricing only, and that is going to burn holes in many wallets. Fable 5 burns roughly twice as many tokens as Opus 4.8, which makes it borderline impossible to use seriously on a standard $20 subscription. If you want real usage, you need at least the $100 plan. If you have been putting off big cybersecurity projects, use your tokens on it within the window before it comes off.
10. How to access Fable 5 (8:13)
Two paths, both simple:
1. Claude desktop application: update the app, and you will get a prompt telling you that you can switch to Fable 5.
2. Claude CLI: update it, and you will get the same prompt saying you have access to Fable 5.
11. My final recommendation (9:19)
Test it while it is on your plan so you know whether it fits your business, but be honest with yourself: about 80% to 90% of businesses out there will be just fine using Opus 4.8, the Opus family, and even Sonnet 4.6. Do not burn a hole in your wallet if you do not need to. I will keep testing it over the coming days and follow up with a tutorial on making the best of it.
Where to go next
See what this model class is actually great at in Build a 3D Learning Game with Claude Fable 5, and if you are optimizing a normal Claude subscription instead, start with How to Choose the Right Claude Model.
