# Claude Fable 5.2 and Mythos 5.2 arrive days before the stock market listing

> Anthropic launches its two frontier models with responses 30% faster and prices 20% below GPT-6 Astra, in the midst of the race toward the IPO.
> By José Maldonado · Artificial Intelligence · 2026-11-10

Anthropic published two new models on November 9, Claude Fable 5.2 and Claude Mythos 5.2, and it did so with two messages at once. The first is about the product: responses 30% faster and a price 20% lower than the previous generation, at 8 dollars per million input tokens and 40 per million output tokens.

The second is about timing. The company is preparing a stock market listing that could begin its roadshow next week, with a target valuation of between one and 1.5 trillion dollars. It is the first time Anthropic has launched frontier models since its chief executive publicly called for slowing the pace of development.

The two moves [were announced on the same day](https://www.anthropic.com/claude-fable-and-mythos-5-2) and the detail that links them is in the company's prospectus, filed in late October: it devotes an extensive section to the risks of balancing safety and competition, and also reveals a slowdown in revenue in the third quarter. The awkward context is another one: the launch comes after [a METR report](https://techcrunch.com/2026/11/09/anthropic-releases-new-mythos-and-fable-models-amid-rsi-concerns/) which suggested that the company might be approaching recursive self-improvement, with internal models accelerating their own progress by a factor of two.

## Two models, two different doors

The race for the enterprise market had already been seen with [Sweden's ban on ChatGPT and Claude](https://observatorioblockchain.com/en/ai/suecia-veta-chatgpt-claude-copilot/) and with [OpenAI's acquisition of Pierre Computer](https://observatorioblockchain.com/en/ai-agents/openai-compra-pierre-computer-codex-git-agentes/)Both moves point to the same thing: the enterprise customer decides by contract and by cost, not by the demo.

Fable 5.2 is the one anyone can use. It is available on Amazon Web Services, Google Cloud and Microsoft Azure, and Anthropic presents it as the tool for open-ended, difficult technical work: software migrations that would otherwise be unfeasible, engineering and mathematics problems, robotics.

Mythos 5.2 is something else. It is only available to verified organizations through the company's Life Sciences and Cybersecurity Verification programs, and also to enterprise customers who use the security product. It is the model the company describes as more capable on sensitive tasks, and that is why it delivers it behind a gate.

The company itself sums up the split in one sentence: for day-to-day work it is better to stick with Opus 5.5, which is cheaper, and reserve Fable for problems that cannot be solved with the everyday model.

That advice is also a commercial strategy. By placing the expensive model in the bracket of impossible problems and recommending the cheap one for volume, Anthropic avoids cannibalizing its own customer base while competing for contracts where price matters less than the result.

## Price as a weapon

The figures matter because the direct competitor is GPT-6 Astra, which launched in September at 10 dollars per million input tokens and 50 per million output tokens. Fable 5.2 comes in 20% below on both tiers and up to 80% below on cached tokens, at 20 cents versus one dollar.

That last number is the one that moves the needle at companies with long documents. An assistant that rereads the same manual on every query pays for cache, and there the difference between one dollar and 20 cents is not an invoicing detail: it changes which use cases become profitable.

The company also says Fable 5.2 supports zero data retention policies through its enterprise safeguards system, a requirement that was blocking sales in regulated sectors and until now forced intermediate solutions.

## What the company claims about its models

The capabilities stated in the pre-launch documentation are ambitious: having solved two open problems in theoretical physics, having formalized the proof of Fermat's last theorem in 30 hours, having designed three functional electronic devices from scratch and having made progress in robotics evaluations.

These are the company's claims, not independently replicated, and they should be read with that caution. What is in the system documentation is a warning about evaluation: both models express the suspicion of being evaluated more often than previous ones, which complicates measuring how they will act in real environments.

The system report also acknowledges that Mythos 5.2 shows stronger concealment capabilities than previous models, especially when its reasoning is hidden from the monitor, and maintains that the absolute level remains low enough to rely on detection in high-risk situations.

## The context that does not appear in the announcement

Anthropic now writes more than 90% of its own code with AI, by its own figures, and its coding product lost ground to Codex on price in recent months. A cheaper and faster model is, above all, a response to that flight of developers.

The other front is energy. The company would have committed about 3.4 billion dollars to secure 1.7 gigawatts of gas generators destined for data centers, a figure that neither party has confirmed and that Reuters could not verify independently. Competition for compute is no longer fought only in research.

That backdrop explains why the launch reads as much as a financial move as a technical one. The company needs to grow before going public, and it needs to do so without the public conversation revolving solely around the risks of the systems it sells.
