Anthropic’s IPO valuation, which was expected to break SpaceX’s record, is no longer on the table. The heyday of AI investments appears to be at an end, at least for now. CEO Dario Amodei’s company had previously promised to roll out improvements to its models gradually. Now come Fable 5.2 and Mythos 5.2, both based on the same, improved underlying model. What should we make of the IPO plans and these new state-of-the-art models? What is the master plan of the leader in AI?
With the unveiling of the Fable/Mythos 5.2 line, Anthropic is taking a rhetorical middle ground. Although the improvement is iterative, the company claims that software migrations previously deemed impossible are now feasible. Three functional devices are said to have been designed by 5.2, and, as has become customary, the model has broken through a mathematical wall. Behind the scenes, the new model managed to formalize Fermat’s Last Theorem within 30 hours.
These are the kinds of achievements that might have served as strong catalysts for an IPO back in June, at the unveiling of Fable 5. But Anthropic’s valuation appears to have been a bubble of its own, with the rest of the AI industry currently taking a beating as well. Setting aside unique issues (Nvidia CEO Jensen Huang’s infidelity cost the company hundreds of billions in market value), the tide has turned for AI expansion, and Anthropic has scheduled its IPO in a different political climate.
LLMs have more to offer (but aren’t showing it)
A unique advantage for Anthropic was that it had a head start of several months. When Mythos first appeared on the public radar earlier this year, the company chose to hold back the model that eventually became Fable 5. After all, Opus 4.5 had already caused quite a stir by the end of 2025: it’s now clear that it was a turning point for AI, where agents truly became agentic and the results were generally reliable, especially in specific areas where the answers were verifiable.
But the next step was Mythos/Fable, a model that, immediately after its full rollout, was deemed too dangerous by the U.S. government to make available to non-Americans. That block was, incidentally, quickly lifted. After OpenAI released GPT-6 Astra and Google more recently released Gemini 4 Argon/Carbon, this new level of AI has become established and well-known. That head start of several months has thus largely evaporated. Unfortunate for Anthropic, just as investors have to decide what it is worth.
Restricting these LLMs means that their increasingly powerful capabilities are channeled toward what is permitted. The formalization of Fermat shows what’s under the hood: 1,000 Fable 5.2 agents took 30 hours, where an earlier, unreleased version of Claude still needed eleven days.
Fable 5.5?
Fable and Mythos 5.2 are, on the whole, iterative when we look at the benchmarks. The performance figures are the best across the board, with GPT-6 Astra, Gemini 4 Carbon, and its own Opus 5.5 serving as points of comparison. Fable has, however, become more affordable: $8 per million input tokens and $40 per million output tokens, a meaningful advantage over the $10 and $50 of its predecessor. Still, many users will continue to be satisfied with the significantly cheaper Opus 5.5, which Pro users can utilize to its full extent within the usage limits, not just half, as is currently the case with Fable.
These version numbers strongly suggest that Anthropic might roll out a Fable 5.5. After all, there is already a 5.5 version of every other variant: Opus, Sonnet, and Haiku. In all cases, the improvement over their predecessors was more than just incremental, and the feedback was extremely positive. By comparison, the step from Fable 5.1 to 5.2 is modest.
It’s unclear what would have been best for Anthropic. An even better model that feels like the next step after Fable 5, named 5.5 or even 6? Greater safety, as promised here, with gradual improvements? Whichever of the two it had been, it would have made little difference right now. It will be a difficult IPO for Anthropic, partly because of the slowdown in improvements, whether self-imposed or not, but mainly because of the expected delay in building AI infrastructure in the U.S. and beyond. For now, the master plan seems to be primarily to get the IPO done before sentiment shifts further.
Read also: Claude Fable 5 is Mythos for the masses