Open your mouth and a skull on your screen opens its jaw. Walk through a game level built from a short brief. These are among the projects people are sharing after Claude Fable 5.2’s release on 9 November.

Could your team build a training game or an interactive lesson this way? Start with one experience you can test. The examples below come from creators’ accounts and Anthropic’s hardware report; AI Boss has not reproduced their builds.

The “X-ray” mirror follows your face

Reddit user Sourcecode12 built a webcam anatomy mirror. It places a 3D skull, brain and facial muscles over the camera picture. The jaw follows your mouth; a hand view places bones over your fingers.

The overlay fits a model from the BodyParts3D dataset to your head. It cannot show your own bones or diagnose a condition; a webcam cannot take an X-ray.

The creator says Fable wrote the code and prepared the 3D models, while they tested each version and requested changes. The first working version reportedly took under an hour. “One chat” included follow-up requests for fitting controls, labels and softer edges.

For a science teacher or museum educator, this suggests an interactive lesson: let learners switch between structures and explain what each does. Have a subject expert check the labels and positioning before teaching with it. A public X-Ray Mirror repository also offers code to inspect. AI Boss has not audited that code.

A playable shooter with an estimated US$15,000 cost

In his first-hand account, Matthew Berman describes a Gears of War-inspired level built with Fable in Unreal Engine, software used to create 3D games. He reports moving through a town, taking cover, fighting computer-controlled enemies and using timed reloading.

The assignment was one playable level. It took a few days, and Berman estimated the computing used would cost about US$15,000 at the pay-per-use rates for software that calls the model. He gives an estimated cost and does not say he paid that amount. He also reports awkward animation. He puts it plainly: Our games aren’t perfectly polished.

Watch Berman’s demonstration to see the project he describes.

For your team, try a smaller original game: a customer-service scenario with a goal, decisions and consequences. Ask a colleague to complete it without coaching. Check whether the choices teach the intended lesson before spending time on elaborate scenery.

People are building places and stories you can explore

Stonehenge in your browser

A creator posting as Miyamoto_-_Musashi reports using Fable 5.2 at High effort to build an interactive Stonehenge scene. The public demonstration offers sunrise, midday and sunset controls, visitors and different viewing angles.

A heritage attraction could try the same format for a guided site introduction. Start with a small, documented area and approved facts. Ask a historian to check the reconstruction; a convincing scene does not establish historical or astronomical accuracy.

A pop-up book made by a team of models

A creator shared a 3D pop-up-book film made with Fable supervising three local assistants running the same Qwen model on a MacBook. Fable planned and checked work; the local assistants wrote the files. The creator reports a US$2.98 cloud bill compared with US$15.77 when Fable worked alone.

Some delegated tasks initially failed. The creator works on Atomic Agent, the tool being promoted. This single comparison does not establish equal quality or the total cost of using the local computer.

For a campaign or lesson, the useful experiment is one short animated scene with approved wording. Request a storyboard first. Check its meaning before asking for motion, narration or more scenes.

The hardware examples include a handheld virtual pet

Anthropic and Diode Computers report three working prototypes: a speakerphone, a voice-assistant hub and a handheld virtual pet. Fable designed circuit boards and cases and wrote the software inside the devices. Diode’s engineers physically built and tested them.

Anthropic says the prototypes worked without fixes to their circuitry. It also states that they had not completed electrical validation, thermal testing or certification. Diode has a commercial relationship with Anthropic. Their report covers prototypes. It gives no approval to sell a finished product.

If you have a product idea, begin with a visual model and a list of requirements. For example, sketch the size, buttons and screen of a desk reminder device. Use that to brief an engineer, who can assess the electronics and testing needed before manufacture.

Concept map: a webcam anatomy overlay, a 3D game, an animated scene and an electronics prototype each need different checks before use.

What the benchmark labels tell you

A demonstration shows one selected project. The benchmark scores below cover specified sets of tasks, with different checks for success. Neither tells you how much work your own project will need.

Three Fable 5.2 results, checked on 11 November 2026
Published resultWhat it measures
AutomationBench 1.0.6: 51.4%338 of 657 simulated business tasks met every required check on the records left behind. This used Max effort and allowed other models to attempt steps that Fable refused.
AA-Briefcase v1.1: 66.1% rubricChecks within professional assignments such as spreadsheets and memos. The rubric is the marking checklist. A separate Elo rating ranks performance relative to other submissions, combining checklist results with comparisons of analysis and presentation. How grading works.
FrontierCode 2.0: 49.1%, XhighA score for software changes against the maintainers’ checklist. Some checks count more than others. Failing a mandatory requirement earns zero. It is not the percentage of office jobs automated. Scoring method.

Effort settings control how much processing the model uses. Xhigh and Max name those settings; both use the same model. Higher settings can take longer and cost more; their benefit depends on the task. Compare the same test version and setup.

Give it a project you can judge

Choose a small result that somebody will use: one lesson, one product view or one game scene. Give Fable the approved content, the audience and the checks it must pass. A nontechnical person can define those requirements even when a developer needs to help with the build.

Try a customer-service training game

Use the practice brief below in a coding workspace that can create and preview a web page. Ask it to show the plan before starting. AI Boss wrote this exercise. It is offered as a brief to try; no Fable test result is presented here.

Build a small browser game for a new shop assistant.
The customer wants a refund for a faulty JMD 8,000 kettle.
Shop rules for this exercise:
- A receipt and purchase within 30 days are required.
- The assistant can approve refunds up to JMD 5,000.
- A manager must approve larger refunds.
- A missing receipt goes to the manager for review.

Create three cases: a valid receipt from 10 days ago;
no receipt; and a receipt from 45 days ago.
Let the learner choose the next action. Explain the rule behind the answer and let them retry.
Use a simple illustrated shop, original shapes and text.
Include keyboard controls, phone controls and restart.
Do not collect customer details or connect to payments.
First show the screen plan and answer key. Then build one case for approval before adding the other two.

Check the answers: The first case requires manager approval because JMD 8,000 exceeds the assistant’s limit. The second needs manager review because the receipt is missing. The third falls outside the stated 30-day rule, so the assistant must explain the rule and cannot approve a refund under it. Test every choice, retry and restart on a phone as well as a computer. Check that the game does not reward a wrong answer as correct.

Measure the work left after the demo

Set a spending limit in your account or project controls where available, and check usage between stages. A request to “stay within budget” is not a spending control. Record preparation, waiting, checking and correction time, along with charges and failed attempts.

Anthropic recommends Opus 5.5 for most everyday work. Try Fable where your usual tool struggles, and compare both against the same acceptance checks. Check current access and the temporary launch allowance before committing to a long build.

After a colleague tests the first scene, ask what confused them and change that part. Add the next scene when the first one works without your explanation.

Sources & further reading

Sourcecode12: webcam X-ray mirror, original Reddit post ↗X-Ray Mirror: public repository ↗Matthew Berman: Fable 5.2 field tests, 9 November 2026 ↗Matthew Berman: video demonstration ↗Miyamoto_-_Musashi: Stonehenge creator account ↗Stonehenge: public demonstration ↗GapNew4766: pop-up-book model comparison and affiliation ↗Anthropic and Diode: electronics prototype report, 9 November 2026 ↗Anthropic: Claude Fable 5.2 and Mythos 5.2 launch, 9 November 2026 ↗Zapier: Fable 5.2 AutomationBench results and setup ↗Artificial Analysis: AA-Briefcase v1.1 tasks and grading ↗Cognition: FrontierCode 2.0 scoring methodology ↗Anthropic: effort settings and model-specific guidance ↗Claude Help: Fable 5.2 launch offer and expiry ↗

Sources checked 2026-11-11. Examples and recommendations are original editorial analysis.