1. What did Anthropic release on Monday?
Anthropic released Claude Fable 5.2 on November 9, along with Mythos 5.2, a restricted sibling model available only to verified organizations. The company promises stronger coding and reasoning than before and lower per-token prices. Early testers report striking results, and also estimated bills in the thousands of dollars for single projects.
Anthropic presents the models as able to turn simple instructions into working software, design physical objects and take on research problems that once required teams of specialists. The cost question arises from how that work gets done: the model does not answer one question and stop, but can run for long stretches on its own.
2. Why can Claude Fable 5.2 be so expensive to use?
The high bills do not come from one costly prompt. Fable can work for days, writing code, testing it and fixing its own errors, and usage charges accumulate throughout. Billing for these models is based on tokens, the small units of text that the model reads and writes, so a project that loops through thousands of steps consumes billions of them.
Developers who pay by usage through the API see these charges directly. Subscribers on fixed-price plans face a different limit, since long autonomous runs can use up their allowances quickly. In both cases, the economics depend on how much work the model is left to do without supervision.
3. What did early testers build with it?
Matthew Berman, an AI-focused YouTuber, called Fable 5.2 the best AI model he has used. In a post on November 9, he said he had used more than 100 billion tokens testing it and described it as the most autonomous model to date. His projects included playable games and a browser-based Excel clone with formulas and pivot tables.
One level inspired by the game Gears of War: E-Day took 43 hours and 48 minutes of build time and processed 8.63 billion tokens. Using standard pay-as-you-go developer pricing, the estimated API cost reached $14,710.
The Excel clone came to about $11,700 after discounts for cached input, which is reusable information the model does not have to process again. A game in the style of Fall Guys used more than 9 billion tokens over 71 hours, for roughly $13,000.
These figures are estimates of what the work would cost at API rates, not necessarily what the testers paid. Berman also reported awkward animations, bugs and features that still needed testing, so the projects were impressive but not finished products.
Not every test is expensive. A simple orange cylinder built as a cheap test cost under $40, roughly two months of a Claude Pro subscription. A physics simulation of a water balloon cost $91.
The comparison shows the spread in what users can spend. The model is aimed at serious work and customers with deep budgets, not routine tasks, and the difference between a $37 tower and a $15,000 video game is mostly the amount of autonomous labor requested.
4. What does Anthropic claim about the new models?
Anthropic says the new models produce output up to 30% faster than their predecessors. It also says Fable helped solve two open theoretical-physics problems about black holes and design three working electronic devices. These are claims from the company, and they call for independent scrutiny before they are treated as established.
On FrontierCode 2.0, an internal Anthropic benchmark that measures the share of proposed coding changes accepted into real open-source projects, Fable 5.2 scored 49.1%, up from 32.4% for Fable 5.1. A benchmark of that kind tests work that resembles professional software development, which helps explain why users are pushing the model well beyond ordinary chat.
Fable and Mythos share an underlying model. Fable carries stricter safeguards against requests involving dangerous biology and cybersecurity, while Mythos is limited to verified organizations working in those fields.
5. What does Fable 5.2 cost per token?
Standard API pricing for Fable 5.2 is $8 per million input tokens and $40 per million output tokens, down from $10 and $50. An optional fast mode generates output up to 2.5 times faster but doubles the rates, to $16 and $80.
Competing models are far cheaper on paper. Kimi 3.1 Pro, regarded as one of the best Chinese AI models, costs $2 per million input tokens and $10 per million output tokens. The latest DeepSeek model costs $0.6 and $1.98 outside peak hours, on the same per-million basis.
By Anthropic's own standards, Fable's per-token price is twice that of Claude Opus 5.5, the model Anthropic recommends for everyday work. Fable is meant for problems that cheaper models struggle to finish, and the lower price per token does not offset the large number of tokens that long autonomous runs consume.
6. How do token-based bills add up?
A token is a small chunk of text, often a word fragment, that a language model reads or writes. Providers charge separately for input tokens, which are what the model is given, and output tokens, which are what it produces. Output tokens cost more because generating them requires more computation, and the Fable 5.2 rates reflect that: $8 per million for input against $40 per million for output.
A long agentic session multiplies both sides of the bill. Each step in which the model reads its earlier work, runs a test and writes a correction adds tokens, and the context it carries forward grows as the project grows. This is why one project can process 8.63 billion tokens, as the Gears of War-style level did, even though no single request looks large.
Cached input softens the effect. When the model repeatedly reuses the same material, discounted rates for cached tokens apply, which is how the Excel clone ended up near $11,700. Anthropic's improved discounts for reused inputs in Fable 5.1 were aimed at exactly this pattern of repeated revisiting.
7. Why does autonomy change the cost calculation?
With earlier chatbots, a user's spending tracked how many questions they asked. With a model that can write, test and repair software by itself for 43 hours or more, spending tracks how long the user lets it run. The user pays for labor rather than for answers.
That shift makes the comparison with human work more relevant. A level of a modern shooter or a spreadsheet application would ordinarily take a team weeks or months, so the bills reported by testers, from $11,700 to $14,710, should be set against the human effort that comparable work would otherwise require. Fable is intended for serious work and well-funded users, and the tester's own notes on bugs and awkward animations show that the output still needs human review. For an individual hobbyist, by contrast, the same totals would be out of reach, which explains the contrast between the $37 tower and the $15,000 video game.
The practical effect is that budget control becomes part of using the model. Setting limits on how long a session may run, or on how many tokens it may consume, is the user's main tool for keeping a long autonomous job inside a planned amount.
8. How should the headline claims be read?
Several of the strongest statements come from Anthropic itself, including the 30% speed increase, the two theoretical-physics problems and the three working electronic devices. Independent verification has not yet been reported for those claims, and the benchmark behind the FrontierCode 2.0 figure is internal to the company.
The testers' results are more concrete but have limits too. They are estimates calculated at API rates, they come from a small group of early users, and they cover projects chosen to show off the model. A score of 49.1% against 32.4% for the previous version is a large improvement on Anthropic's own measure, yet it also means that around half of the proposed changes would not have been accepted.
The new launch offers give subscribers a chance to try this for themselves, within the caps and deadlines described above. Taken together, the evidence points to a model that is more capable and cheaper per token than its predecessor, but whose total cost for ambitious projects can be much higher than for ordinary use. Readers weighing it against cheaper alternatives such as Kimi 3.1 Pro or the latest DeepSeek model should compare finished results rather than price lists alone.
9. How have earlier Fable releases handled costs and rumors?
Fable 5 reportedly burned through some subscribers' quotas within minutes at launch. When Fable 5.1 arrived in September, Anthropic kept premium pricing but improved discounts for reused inputs, which reduces the cost of long sessions that repeatedly revisit the same material.
Rumors of a Fable 5.5 circulated last week, but the actual release is version 5.2. It comes with cheaper token rates, although its capabilities still encourage the kind of extended work that produces large bills.
10. Who gets access under the launch offer, and until when?
Pro users and standard Team seats receive one week of Fable access, capped at half their weekly allowance, along with a usage reset. Max, Ultra and eligible premium seats receive between one and three resets, depending on their subscription history.
The promotional access for Pro users ends on November 16, after which Fable uses paid credits on those plans. Resets for eligible Max and Ultra subscribers last until March 8, 2027.
Anyone using the free week or the resets for heavy workloads should watch their consumption, since a single large project can exhaust an allowance quickly. Berman's $14,710 Gears of War example is a reminder of how large the totals can become when the model is allowed to work unattended.