Fable 5.2 first user tests: games and video fast, cost in doubt

On November 10 and 11, the Chinese outlet QbitAI collected the first hands-on results from users of Claude Fable 5.2, Anthropic's new model, under a headline that sums up the mood: "Opus suddenly isn't so appealing." The cases come mostly from developers and researchers on the social network X. They are demonstrations by internet users, not official benchmarks, and none has been independently verified, so they should be read as ideas for what to try, not as performance measurements.

What users built with Fable 5.2

According to QbitAI, which relays posts from X, one user burned through about 400 billion tokens to produce a whole spread of games, videos and interactive worlds. Wharton professor Ethan Mollick mobilized about 25,000 agents to generate more than 450 3D worlds created in real time by code. Another user fed the model 500-year-old Leonardo da Vinci manuscripts and got back 3D models of his machines that can be handled and moved.

Games are the largest category. QbitAI lists several examples:

• The entire Earth turned into a playable game map, with weather synchronized in real time (a demo shared by Matt Shumer).
• A 3D portal shooter that its developers reportedly already used for team play sessions at the office (a post by Sholto Douglas).
• A Call of Duty style shooter, with 3D modeling, textures and sound, built in 8 hours with no human feedback (a post by Felix Rieseberg).
• An underwater version of the "big fish eats small fish" game, delivered in 76 minutes for about 280 yuan, which its author called the best version he has made.

Beyond games, QbitAI mentions a piano sonata that was composed, synthesized and paired with a video where the notes follow the keys; a short film generated purely from code about the big AI moments of 2026; a promotional video for a SaaS product made in 10 minutes from a single prompt; and a complete 3D room built in Blender, with 13 camera cuts, 4K rendering, music and sound effects, delivered in 52 minutes for about 38 yuan in rendering costs. All of these figures are self-reported by the users.

What the official figures say

According to QbitAI and the Taiwanese outlet iThome, both citing Anthropic's materials, Fable 5.2 generates output up to 30% faster than Fable 5.1, supports a context of one million tokens and up to 128,000 output tokens. Its score on the software engineering test reaches 69.9%, 13.6 points above the previous generation. API prices fall by 20%: 8 dollars per million input tokens and 40 dollars per million output tokens. iThome adds a fast mode, up to 2.5 times quicker, priced at 16 and 80 dollars per million tokens, cache reads at 0.20 dollars per million and half price for batch processing.

iThome also explains that Fable 5.2 and Mythos 5.2 share the same underlying model. They differ in safety protections and eligibility: Fable is open to the general public, while Mythos is reserved for organizations admitted to Anthropic's verification programs for cybersecurity or life sciences.

The most interesting point is cost efficiency. According to iThome, citing Anthropic's benchmarks, Fable 5.2 scores 45.3% at the High effort setting of FrontierCode 2.0, against 45.5% for Opus 5.5, but at roughly half the cost per task; at Xhigh it rises to 49.1%, the best score among the released models Anthropic has tested. The company admits it is not ahead everywhere: Gemini 4 Carbon leads on a business workflow test, and Opus 5.5 on a knowledge work benchmark. QbitAI writes that the model "beats Opus 5.5's 45.5%," which differs slightly from iThome's High figure; it may refer to a higher effort level, but the article does not say.

The bill and the quotas are the real friction

QbitAI also reports the other side of feedback from the developer community: the model is powerful, but it burns many tokens, especially at the maximum reasoning level, and cost per task and quota use climb quickly. Some complain that on several tasks the gain over Opus 5.5 is limited while the price is said to be twice as high. These are user testimonies, with no verifiable data in the article.

For most coding and knowledge tasks, we still recommend starting with Claude Opus 5.5, our strongest daily driver.

In other words, "Opus is no longer appealing" reads more like a social media headline than a conclusion from Anthropic: the maker keeps Opus 5.5 as the first choice for everyday work and reserves Fable 5.2 for harder, longer tasks that are worth paying more for.

What this changes for a Chinese company

For a Chinese company thinking of evaluating Fable 5.2, three points stand out. First, do not choose on the strength of demos: test the model on your own hardest real tasks and record the total cost and token use per task, not just the price per million tokens, because the bill for the same model can vary several-fold depending on the reasoning level. Second, treat Fable 5.2 as a heavy tool for difficult cases, and keep cheaper models, including Chinese ones, for daily work such as code completion, document tidying or customer service, routing by difficulty. Finally, none of the sources cited here says whether the model can be used directly from mainland China: purchasing route, data export and compliance need to be checked with the cloud provider and the legal team before launching a pilot.