Claude’s 100B-token coding spree turned AI autonomy into a $14,710 API bill 🧾

Anthropic’s Claude Fable 5.2 and Mythos 5.2 can autonomously write and test software for days, but one 43-hour project consumed 8.63 billion tokens and an estimated $14,710 in API costs. The release arrived on Monday, with Mythos 5.0 also identified as a model name in the announcement, and generated output up to 30% faster than its predecessors.

AI YouTuber Matthew Berman described Fable 5.2 as the “best AI model” he has used after creating playable video games and a browser-based version of Microsoft Excel with formulas and pivot tables. A Gears of War: E-Day-inspired level required 43 hours and 48 minutes of build time and processed 8.63 billion tokens, pushing the estimated pay-as-you-go API cost to $14,710. The Excel clone cost an estimated $11,700 after discounts for cached inputs, while a Fall Guys-style game consumed more than 9 billion tokens over 71 hours and cost roughly $13,000.

In a November 9, 2026 post, Berman wrote: “I've burned over 100 billion tokens early testing Fable 5.2. It is BY FAR the best model I've ever used. Here's my full review (including Gears of War clone): > Fable 5.2 is the most autonomous model to date. Given a task with little direction, it makes… pic.twitter.com/0xTzqTY040 — Matthew Berman (@MatthewBerman).” The projects were not flawless, with Berman reporting awkward animations, bugs and features that still required testing. Less expensive experiments included a cute orange cylinder built for less than $40, equivalent to two months of Claude Pro, and water-balloon physics simulation produced for $91.

Anthropic said Fable helped solve two open theoretical-physics problems involving black holes and designed three functioning electronic devices. The company’s internal FrontierCode 2.0 evaluation, which measures the percentage of proposed coding changes that would be accepted into real open-source projects, placed Fable 5.2 at 49.1%, compared with 32.4% for Fable 5.1.

Fable and Mythos share the same underlying model, but Fable has stricter safeguards around danger. Anthropic did not position the models for routine tasks, instead presenting them for complex software development, physical design and research requiring sustained autonomy.