Claude Fable 5.2 has been Anthropic's most capable model since November 2026. Three things change for a company. Developer pricing drops by around 20 percent. Responses arrive up to 30 percent faster than on Fable 5.1. And the model takes a leap on anything involving geometry, technical drawing and chart reading.
To see what that means outside published tests, I gave the model a concrete case. A two-dimensional floor plan for a food service unit, 126 square metres, 9 metres of frontage. It returned a three-dimensional model you can walk through, split into five zones, with four layouts to compare. Here is what I take away about the model, its price and its limits.
What changes with Claude Fable 5.2?
Claude Fable 5.2 replaces Fable 5.1 on hard technical work. Anthropic shipped the model in November 2026, alongside Claude Mythos 5.2, which runs on the same base but stays restricted to organisations verified for cybersecurity and life sciences. Five changes matter for a company.
- Pricing drops. Around 20 percent less per million tokens, in both directions.
- Speed goes up. Outputs arrive up to 30 percent faster than with Fable 5.1.
- A fast mode appears. It runs up to two and a half times faster, at double the price. It is available on Claude Code, on the Claude platform and at two cloud hosts.
- The model can delegate. Two roles arrive. As an advisor, a smaller model consults Fable on a hard point. As an orchestrator, Fable hands work out to faster models.
- It refuses less often. Anthropic reports around 55 percent fewer blocks than Fable 5.1, on legitimate requests the previous model turned down.
The model identifier is claude-fable-5-2. It is available on every paid plan, through the API and at the three major cloud hosts, according to Anthropic's announcement.
How much does Claude Fable 5.2 cost?
Claude Fable 5.2 pricing is read per million tokens, and it depends on the mode. The figures below are converted into euros from the price list Anthropic published in November 2026. The cut applies to usage-based pricing. Consumer subscriptions have not moved.
| Mode | Input price, approximate | Output price, approximate |
|---|---|---|
| Standard | 7 euros | 37 euros |
| Fast | 15 euros | 74 euros |
| Batch, asynchronous | 4 euros | 18 euros |
| Fable 5.1, for reference | 9 euros | 46 euros |
Those figures cover the API. For a business leader who simply wants to use Claude in a browser, the question is the subscription, and we covered it in our article on Claude Max pricing. Fable 5.2 is included there, up to half the weekly quota on Pro plans and standard Team seats.
Fable 5.2 against Opus 5.5, GPT-6 Astra and Gemini 4
Anthropic publishes nine comparative tests with its announcement. Claude Fable 5.2 comes first on eight of them. The ninth is the most interesting one for a company, because it measures the ability to run a business workflow end to end, and Gemini 4 Carbon stays ahead there with 53.9 percent against 51.4 percent.
A benchmark a vendor publishes about its own model deserves caution. This one still has a virtue: it shows where the model loses. What I take from it is the gap on computer-aided design and on chart reading, two areas where language models were mediocre until now.
I gave Fable 5.2 a 2D plan, here is what came back
The test covers the layout of a food service unit. Starting point, a fictional plan of 126 square metres, 9 metres of frontage and 14 metres of depth, drawn for a demo. No imported 3D model, no reference image, no furniture library.
The model produced four deliverables from that single plan. A model split into five zones, an eye-level walkthrough with eight stops, a 6.5 metre counter with a 40 seat dining room, and four layouts to compare. All of it fits in around 440 lines of code.
I published the full walkthrough video on LinkedIn in November 2026, with the detail of all four layouts and the decor borrowed from the chain.
Why the model goes through code and not through 3D software
Claude Fable 5.2 writes the program that draws the restaurant, and that program then runs in a browser. No image is produced at any point. The difference is decisive for professional use.
When a model generates an image, every new attempt is a roll of the dice. You rerun it, you get something else, you pick the least bad one. When it writes code, you fix one line and the rest stays put. Moving the counter by 50 centimetres becomes a simple edit to the program.
That is the point I made in my newsletter on code in August 2026. A deck, a document, an edited video, a scale model: behind every deliverable there is a structure, so there is code, so there is something you can delegate. To see how that kind of work gets steered without being a developer, our Claude Code guide lays out the method.
Can other models do 3D too?
Every recent frontier model can produce a three-dimensional scene from a description. Code-written 3D is shared ground. I filmed the case of Gemini 3 and its browser-playable 3D games, a good illustration of what the previous generation already did.
The gap shows up over long runs. Ethan Mollick, a professor at Wharton, asked Fable 5.2 for a series of 3D worlds linked by portals. He got 204 of them in about twenty hours, then more than 450, each drawn live by code. He also notes the limit in his write-up: quality is uneven, and the agents never noticed on their own that their output was turning repetitive.
What this changes for a small or mid-sized company
The value for a company lies in the trade-off the model makes visible before any money is committed. On my test the four layouts give very different answers to the same question, how many covers for how much waiting.
A chain opening a new outlet arbitrates between seat count and service flow, and it does so before signing the lease. The same mechanics apply to a store planogram, a workshop rearrangement or a trade show stand. We detail that kind of use on our AI for retail page. For day-to-day visual production, Claude Design remains the easier tool to put in everyone's hands.
The limits of Claude Fable 5.2
The first version of my model was unusable. Overexposed image, a camera that went through a plant, then through a wall. It took several correction passes to get to something presentable. The model produces fast. Getting it right takes several rounds.
- The model does not judge its own work. That is Mollick's finding across his series of worlds, and I found it on a single unit. Someone has to look and decide.
- A scale model is not a construction drawing. It replaces neither the architect, nor accessibility rules, nor fire safety.
- The figures it shows come from your assumptions. The model computes what you ask it to compute. A wrong aisle width going in gives a wrong cover count coming out.
- Agent mode gets expensive fast. A model working alone for an hour burns a lot of tokens, and the bill is read per million.
Should you move to Claude Fable 5.2 now?
Yes for hard technical work, no as a default model. Fable 5.2 earns its price when the output is long, structured and checkable: a program, a scale model, a code migration, a numbers-heavy analysis. For everyday writing or document summaries, Opus 5.5 does the job for about half the cost.
The real question for a small or mid-sized company already sits beyond the choice of model. It is knowing which tasks deserve the heavy artillery, and who on the team can frame the request. Tandem works with more than 50 companies on exactly that, and our Claude training starts from there.
