How Much Does Fable 5.2 Cost? Real Token Bills by Project

How much does it cost to run big projects on Fable 5.2?

It depends almost entirely on how long the model runs. In two creator walkthroughs, short Fable 5.2 builds (a Blender house, a 3D-printable tower, an explainer animation) came to roughly $25 to $91 in API-equivalent cost. Long Ultra Code runs (a Gears of War clone, a Fall Guys clone, an Excel clone) came to roughly $11,700 to $15,000. "API-equivalent" means what the usage would cost at pay-per-token API rates, not necessarily what anyone paid on a subscription.

What did each project actually cost?

Here are the figures the creators gave, with the cost per hour worked out from their stated build times. The per-hour column is my arithmetic, not a number from the source.

Two smaller notes. In the other review, one demo (a one-shot simulation) was estimated at 30 to 50 cents in tokens. And the creator gave no total for the 3D European town or the DaVinci Resolve-style video editor, so there is no number to quote there.

Why do the big projects cost so much more per hour?

Token volume. The small builds used roughly 10 to 40 million tokens. The big ones used billions. By my math, the water balloon simulation burned around 16 million tokens per hour, while the Fall Guys clone averaged on the order of 125 million per hour.

Most of that volume was input read from cache. A long-running agent keeps rereading the project it has built so far: the code, the plan, its earlier decisions. Cached reads are billed at a discount, which is why the cost per million tokens in the large runs (about $1.45 by my arithmetic for both the Excel and Fall Guys projects) was lower than in the small ones (about $2.20 to $3.70). Even at that discount, billions of tokens add up fast.

The practical lesson: the price tag tracks how many hours the agent works, and how much context it carries, more than how impressive the result looks.

What is Ultra Code mode?

The review says only that the creator "put it on Ultra Code mode and just let it go" for the Gears of War clone. The sources do not explain what Ultra Code does internally, and they do not describe sub-agents, so I can't tell you how it splits work.

What the numbers do show is that Ultra Code runs are long and unattended. A three-line prompt ("build the most accurate clone of Gears of War Eday using Unreal Engine… only one level") turned into a multi-day run and a roughly $15,000 API-equivalent bill.

As general background (not a claim about Ultra Code specifically), tools that split a job across several AI agents tend to multiply token use, because each agent needs its own copy of the context. If you try a mode like this, assume the bill scales with how many things run in parallel and how long they run.

When is Ultra Code worth it?

Run the math against what you would pay a person or a vendor.

• Worth considering: a deliverable you would otherwise pay a team weeks to build, where "good enough and working" is the bar. The creator's Excel clone took almost 50 hours of machine time and was described as having no missing features he could find. If an equivalent human project costs more than the bill, the numbers can work.
• Probably not worth it: exploratory work, anything with a vague goal, or anything you could do in a short run. The $25 to $91 projects covered a lot of ground.
• Never worth it unattended without a ceiling: an open-ended prompt with no limit on scope. The Gears of War run was still going when the creator quoted its cost.

Also remember these were demos by creators with a promotional interest. Some outputs had flaws the creators admitted: an awkward reach-over-barrier animation in the game, a tree in the middle of the Blender driveway, and a Roadster recreation the creator called "the first bad one."

How does the free usage reset work?

According to the first review, when you start a new chat with Fable 5.2 and ask it what model it is, Anthropic shows a notification in the top right corner saying you get a free usage reset. The creator recommended using it to test the model.

The sources don't say how long the reset lasts, which plans get it, or how much usage it restores. Check the notification in your own account for the terms. The sensible use: spend it on a small, well-defined project, ideally one in the $25 to $100 range from the table above, so you learn what the model does for your kind of work before paying for it.

How good is Fable 5.2 compared to Opus 5.5?

Better, but not dramatically. The review cited these benchmark scores:

• Agentic Coding Frontier Code 2.0: 45.5 for Opus 5.5, 49.1 for Fable 5.2.
• Agentic Coding (SWE-style) benchmark: 62.3 versus 69.9.

The reviewer's own summary was that this is about a 10% bump on coding tests and 15 to 20% on scientific research and business workflows. His advice was to treat it as a solid upgrade, not a sign that everything has changed. A Mythos 2 model launched alongside it, but the reviewer noted it is aimed at people with cybersecurity backgrounds and wasn't what regular users get.

What does this mean for you?

If you're a manager or budget owner: ask for a token estimate before approving AI-built projects, the same way you would ask for hours on a contract. Tens of dollars and tens of thousands of dollars are the same tool; the difference is scope and runtime.

If you're an individual contributor: start with short, specific tasks. A one-hour job with clear constraints (the 3D tower prompt listed printer, nozzle, material and weight limit) cost $37. Specific prompts finish sooner.

If you're considering a long autonomous run:

1. Write down what "done" means before you start.
2. Test the idea at small scale first, using the free reset if you have it.
3. Check usage while it runs, not after. If your plan or tool offers spending limits, set one.
4. Compare the bill to the cost of the human alternative, including the time to review the result.

Could you build your own AI cost tracker?

Yes, and it's a good small project. A simple internal app with a form for project name, model, tokens, hours and estimated cost, plus a dashboard totaling spend by project, would answer the question this article is about for your own team. You could describe it in plain English to Remy, which builds and deploys full-stack apps from a description, instead of keeping the numbers in a spreadsheet nobody updates. Remy bills AI tokens at cost, and its pricing FAQ puts a typical idea-to-deployed-app build at $100 to $500 in tokens, so apply the same budgeting habits to it.