DEV Community

Cover image for I Built a 3D Farm Mini-Game with claudefable5.2 Until Thinking Used Up the Output Budget
boluo
boluo

Posted on

I Built a 3D Farm Mini-Game with claudefable5.2 Until Thinking Used Up the Output Budget

I Built a 3D Farm Mini-Game with claudefable5.2 Until Thinking Used Up the Output Budget

Disclosure (Dev.to): this post is **AI-Assisted. The test and conclusions are mine; I used AI help to draft and then edited the wording.

Anthropic released claudefable5.2 on 9 November 2026. I wanted a call path I could reuse, and a 3D farm I could actually open in a browser. Official materials put this model on interactive 3D worlds and harder game-development work. I used it to open a farm, then kept going until later writes stopped returning code.

I finished the first version the same day, 19 November 2026. The game is called Sunshine Farm. After the playable parts were in, I didn't go back and polish it. What follows is the model, the endpoint I used, the output-budget failure that showed up on later writes, why I didn't use GPT-6 Astra, what official materials change from Fable 5.1, and how this first version became something I could open in public.

claudefable5.2, the release date, and the official brief

claudefable5.2 is the current release in Anthropic's Claude Fable line. The API id is claude-fable-5-2. Officially, it is aimed at hard reasoning work and at agent jobs that have to run for a long time. For everyday use, this is the Fable model you pick when the technical work sits at the hard end.

These official facts are the premises I started from:

  1. It shipped on 9 November 2026. The previous generation, Fable 5.1, shipped on 1 September 2026.
  2. The context window is 1M tokens. Maximum output is 128k tokens.
  3. Adaptive thinking stays on. You cannot turn it off. Default effort is high.
  4. The reliable knowledge cutoff is June 2026.
  5. There is an optional fast mode. Officially, output speed can reach 2.5×.

Official capabilities, as they are written

The list below is official wording. It is not a score I gave the farm after walking it.

  1. Coding is framed around large software work where the goal is not fully specified. Official materials emphasize self-verification, and a greater willingness to question a conclusion that already exists.
  2. For agents, official materials describe it as their strongest lead model at the moment. The claim is that it can keep a loosely defined engineering job running for a long time, and coordinate a large set of sub-agents.
  3. For 3D, hardware, and dynamic work, official pages are blunt. They say it can build interactive 3D worlds. They also list simulation, animation, and code-driven work with sound. The same pages say it stands out when it builds higher-quality 3D scenes from scratch, and on harder game-development tasks. I picked a farm because that is the official emphasis I wanted to test.
  4. Official materials also place it on long-running work in scientific research, financial analysis, data analysis, and machine learning. Image and PDF reading are supported.

I needed those official notes in one place before I called it. The farm was how I tested what would actually come back.

The endpoint I used to call it

I called it through Crazyrouter. The model name is claude-fable-5-2. The endpoint is:

https://api.crazyrouter.com/v1/messages
Enter fullscreen mode Exit fullscreen mode

I checked the model in the Anthropic models section. If you are wiring an editor or coding agent to the same endpoint, the AI coding tool integration notes cover that setup. This run started from the claudefable5.2 page on Crazyrouter.

Later writes failed when thinking filled the output budget

Official docs also flag limits that matter once you start a long writing session. The one that stopped me was the output budget.

  1. Adaptive thinking cannot be turned off. There is no way to disable thinking to save budget. Effort is the only knob for how deep it goes.
  2. One response is capped at 128k tokens, and thinking counts against that cap. On later writes, the model spent the cap on thinking, and no follow-up code came out. That is the limit I hit most directly.
  3. Default data retention is 30 days, for safety monitoring. Zero retention is a separate authorization, not a toggle you turn off in the request.
  4. A required-tool constraint is not supported. That request form returns an error.
  5. Safety policy may block cybersecurity or biology questions, or route them to another model. The farm never triggered this. Official notes still list it.
  6. To continue the same thinking content, leave earlier messages alone. Official prefix checks inspect that prefix. Edits mid-stream can fail the call.

128k is not a small cap. Once always-on thinking fills it, a follow-up that should have added playable code returns nothing you can paste in. That is the point at which this first version stopped being a single model pass.

Why I didn't use GPT-6 Astra for this test

GPT-6 Astra is OpenAI's flagship model. Officially, it is placed on complex coding and professional work. Coding, computer use, and research all sit in its range. I am not saying it cannot make a mini-game. This article is not a head-to-head with claudefable5.2.

I left it aside because the job matched the official write-up for claudefable5.2 more closely. Official materials say it can build interactive 3D worlds. They also mention building 3D scenes from scratch, and harder game development. I wanted to call that model, open a 3D farm you can walk into, and see how far a first version would get.

Official differences from Fable 5.1, as published

If you already call Fable 5.1, the first thing to check is the window. Fable 5.1 shipped on 1 September 2026. Officially it is now Legacy, and it is still callable. claudefable5.2 did not enlarge the context or the output cap: both remain 1M / 128k, with thinking always on, and a June 2026 knowledge cutoff. I would not budget a larger window just because the id changed.

What official materials do claim is a capability and speed shift, not a spec bump.

  1. They say coding, reasoning, and vision improved, and that testers found the newer model clearer.
  2. On official benchmarks, FrontierCode 2.0 at the xhigh tier is 49.1% against 32.4%, and FrontierSWE v2 is 69.9% against 56.3%. Those figures are official and self-reported. This farm test did not re-run Fable 5.1.
  3. Official materials also say generation can be up to 30% faster than Fable 5.1, and they added an optional fast mode for claudefable5.2.
  4. The API surface moved too: prefix checks, whether thinking blocks can be read by other models, and the computer-use tool version. A migration has to take those together.

I have no Fable 5.1 farm to compare. The notes below are from this claudefable5.2 run only.

What the first writes already put on the page

The brief was fixed up front. The game had to open in a browser, in third person, and be walkable. The fields had to be large enough that they would not read as a pocket garden. Farming had to work, with chickens, ducks, and cows on the map. A textile mill, a flour mill, and a dairy had to be reachable, and you had to be able to process materials there.

Early concept frames were pretty, and that was the problem. The plot was too small, too clean, and too uniform. I opened the brief, then let claudefable5.2 do the actual writing.

The first useful return was structural. The page came out. The character came out. Most of the scene came out. Opening it, I saw a straw-hat worker on the field paths and a wooden house already in place. That was close enough that I relaxed. The pattern was obvious on that pass: people and large scenes can land early, and they look like a place.

I only judged the rest after I could walk it.

The fields were the first relief. Hitting start and stepping in, the ground actually went on. Carrots, cabbage, tomatoes, wheat, and cotton were all planted. Some beds were freshly turned. Some were already up. It felt much closer to the brief than those early frames. Scale is the part of this run that held. The loop itself also runs: till, sow, water, harvest, with E to interact and 1 to 5 to pick a seed. It isn't tight. It's a complete pass, not a polished one. There is no save. Walk far and you navigate by memory. I left those issues on this version.

Entering the fields in Sunshine Farm, where the crops are already spread out

The animal area is what stopped the map from reading as one big vegetable plot. A few chickens moved in the coop yard, against a red house and a wooden fence. Standing there felt good. The duck pond has water and reeds, with ducks on the surface, and it has more life than bare grass. The dairy pasture comes later: a few cows inside the fence, and the map finally feels occupied. Scene work still holds. That is still this model's strength.

In the chicken coop yard, with chickens walking inside the fence

At the duck pond, with ducks floating on the water

In the dairy pasture, with a few cows close by

Factories sit on the far side. At the textile mill door, a prompt tells you to press E: 2 cotton for 1 cloth. The flour mill is 2 wheat for 1 flour. The dairy is 1 milk for 1 butter. Processing works. Finding those buildings the first time did not. After the map opened up, I still hesitated and had to learn the route myself. Scale arrived. Wayfinding did not travel with it.

At the textile mill door, where cotton can be woven into cloth by pressing E

Those loops connecting was a relief. They were not a one-pass result I could play immediately. That work was finished later.

How the first version became publicly playable after that

The stall came on the later writes. Thinking used up the output budget, and no code came out. Farming, animals, and factory play had to be finished afterward, or the game would have stayed an empty shell I could not walk for a test. I needed those systems in before I could put a public link on this first version. Once they were in, I stopped. I did not come back to polish this build.

A browser can open it. You can walk, plant, feed, and process. That is enough to inspect this first version. A fuller cycle is not there yet, and I did not start a second version.

The public build still shows the gaps. Scale is in. Signs are not. There is no save. Pointer lock can drop out. Even so, it got far enough that anyone can open it in a browser.

What was still short when I left this version

When I used claudefable5.2 to build this 3D farm, these were the shortcomings:

  1. The playable layer was not completed in one pass. Farming and the factories were finished later.
  2. On later writes, thinking used up the output budget, and no code came out.
  3. The scale opened up, but wayfinding and signage were not built with it. When I walked around, I still had to find my own way.
  4. This first version is still rough. There is no save. Pointer lock sometimes fails, and I had to fall back to drag-to-orbit.

People and large scenes landed first, and the scale opened up. Official materials say it can build interactive 3D worlds. In this run, that claim held at the scene layer. I finished this first version on 19 November 2026, the same day I started it. The spend was $11. It was already good to have come this far.

You can open this first version in a browser: Sunshine Farm. Click start. Move with WASD or the arrow keys, look with the mouse, jump with Space, interact with E or a left click, and pick seeds with 1 to 5.

References

anthropic.com/claude-fable-and-mythos-5-2

anthropic.com/claude/fable

platform.claude.com/docs/en/models/fable-5-2/overview

openai.com/index/gpt-6-astra

Top comments (0)