Fable 5.2 chess 24hrs: 3542 Elo, written from scratch by an AI in 24 hours

[Steve Maughan]

How much did you pay to get the chess engine?

I asked google
how much do you need to pay in order to use Fable5.2 to work for 24 hours to build chess engine

AI replied:

Running Claude Fable 5.2 continuously for 24 hours to build a chess engine will likely cost between $300 and $900+ via the API, as standard consumer subscriptions enforce strict usage caps that will pause a marathon session.
Because Fable 5.2 is billed by token usage rather than by the hour, your final bill depends heavily on how aggressively your coding agent loops, tests, and refactors code.

Edit:After asking a different question the claim is that the price is clearly smaller and I get replies that using claud pro is only 20$ for month and you can use tokens to do the task cheaper so I am not sure about the price.

I am also not sure if using claud pro is the best AI that we have to write chess applications.
I use gemini and not claud pro and I would like to know if generally claud pro is better for writing chess applications.

Hi Uri,

The key point is that Claude Fable 5.2 doesn't run continuously. Much of the time it sits idle, waiting for tasks and sub-agents to finish. A large share of the compute goes into generating games for training data, and there are numerous test runs along the way. As a result, a 24-hour session doesn't come close to my subscription limit. I'm on the $200 plan and use perhaps 10% of it per 24-hour run, and that's with Fable, which is the most token-hungry of the models.

On Claude Pro versus Gemini, "Claude" actually covers a whole family of models, each with different abilities when it comes to building chess engines, and you can see the results for each one on the GitHub page. I don't think Gemini is that strong at this, although Google has just released its latest model, Carbon, which may do better. I haven't tested it myself, but by all means clone the benchmark repository, let it run for 24 hours, and see what it produces.

At the moment I don't think anything can touch Fable 5.2. Opus 5.5 isn't far behind, but Claude is clearly leading the field (IMHO).

Steve

[Bart Weststrate]

How much did you pay to get the chess engine?

I asked google
how much do you need to pay in order to use Fable5.2 to work for 24 hours to build chess engine

AI replied:

Running Claude Fable 5.2 continuously for 24 hours to build a chess engine will likely cost between $300 and $900+ via the API, as standard consumer subscriptions enforce strict usage caps that will pause a marathon session.
Because Fable 5.2 is billed by token usage rather than by the hour, your final bill depends heavily on how aggressively your coding agent loops, tests, and refactors code.

Edit:After asking a different question the claim is that the price is clearly smaller and I get replies that using claud pro is only 20$ for month and you can use tokens to do the task cheaper so I am not sure about the price.

I am also not sure if using claud pro is the best AI that we have to write chess applications.
I use gemini and not claud pro and I would like to know if generally claud pro is better for writing chess applications.

Hi Uri,

The key point is that Claude Fable 5.2 doesn't run continuously. Much of the time it sits idle, waiting for tasks and sub-agents to finish. A large share of the compute goes into generating games for training data, and there are numerous test runs along the way. As a result, a 24-hour session doesn't come close to my subscription limit. I'm on the $200 plan and use perhaps 10% of it per 24-hour run, and that's with Fable, which is the most token-hungry of the models.

On Claude Pro versus Gemini, "Claude" actually covers a whole family of models, each with different abilities when it comes to building chess engines, and you can see the results for each one on the GitHub page. I don't think Gemini is that strong at this, although Google has just released its latest model, Carbon, which may do better. I haven't tested it myself, but by all means clone the benchmark repository, let it run for 24 hours, and see what it produces.

At the moment I don't think anything can touch Fable 5.2. Opus 5.5 isn't far behind, but Claude is clearly leading the field (IMHO).

Steve

Interesting result:

I adapted the Fable 5.2 net without tuning of search parameters to my own search with this result until now:

# Name Points % Elo +/- W-L-D 1 2
----------------------------------------------------------------------------------
1 Selate_BS_small_fable 224.5/343 65.5 +111.0 20.8 +114-8=221 65.5%
2 fable_5_2 118.5/343 34.5 -- -- +8-114=221 34.5%

343/20000 games, 64% draws, time control:30+0.3, book:avt-2026b.pgn

So elo could be much higher. It's a strong small net but the heavily tuned search is not better than another not tuned search to this net.

BUT: Selate_BS_small_fable may be 100 elo better but is's not a 3600+ engine because it is 150 elo weaker than my 3500+ big net engine. And that elo against 1 opponent is inflated and highly unreliable.

Conclusion possible is that this should be tested by a gauntlet, not only against Stash.

[Steve Maughan]

<SNIP>
Interesting result:

I adapted the Fable 5.2 net without tuning of search parameters to my own search with this result until now:

# Name Points % Elo +/- W-L-D 1 2
----------------------------------------------------------------------------------
1 Selate_BS_small_fable 224.5/343 65.5 +111.0 20.8 +114-8=221 65.5%
2 fable_5_2 118.5/343 34.5 -- -- +8-114=221 34.5%

343/20000 games, 64% draws, time control:30+0.3, book:avt-2026b.pgn

So elo could be much higher. It's a strong small net but the heavily tuned search is not better than another not tuned search to this net.

BUT: Selate_BS_small_fable may be 100 elo better but is's not a 3600+ engine because it is 150 elo weaker than my 3500+ big net engine. And that elo against 1 opponent is inflated and highly unreliable.

Conclusion possible is that this should be tested by a gauntlet, not only against Stash.

Interesting experiment!

To clarify, it was tested by a gauntlet of engines. You can find out more in the benchmark repository, that also contains the results:

https://github.com/stevemaughan/chess-engine-benchmark

— Steve

[scchess]

Great works! May I study your orignal prompts? I want to try something similar on Fable too.

[Michel]

This is great of course.

Needless to point out that the LLM has seen the code of > 3500 engines during its training. It would not have come up by itself with all the CPW concepts it relies on. These took many years to develop.

[chrisw]

<SNIP>
Interesting result:

I adapted the Fable 5.2 net without tuning of search parameters to my own search with this result until now:

# Name Points % Elo +/- W-L-D 1 2
----------------------------------------------------------------------------------
1 Selate_BS_small_fable 224.5/343 65.5 +111.0 20.8 +114-8=221 65.5%
2 fable_5_2 118.5/343 34.5 -- -- +8-114=221 34.5%

343/20000 games, 64% draws, time control:30+0.3, book:avt-2026b.pgn

So elo could be much higher. It's a strong small net but the heavily tuned search is not better than another not tuned search to this net.

BUT: Selate_BS_small_fable may be 100 elo better but is's not a 3600+ engine because it is 150 elo weaker than my 3500+ big net engine. And that elo against 1 opponent is inflated and highly unreliable.

Conclusion possible is that this should be tested by a gauntlet, not only against Stash.

Interesting experiment!

To clarify, it was tested by a gauntlet of engines. You can find out more in the benchmark repository, that also contains the results:

https://github.com/stevemaughan/chess-engine-benchmark

— Steve

Blockage to scaling is the 24 hour limit on a) generating training positions and b) time to train.

I find this an interesting experiment, but we reach the end of the line and the pipeline is clear and understood.

Way more interesting is LLM’s playing moves because model. They’re getting kind of quite good, and the process is much more akin to how a human learns and plays, errors and oversights included. When an LLM based on the world’s texts is decisively defeating top GMs, we can consider that quite a milestone.
If such an LLM can defeat SF, which is plausible in time, that’ll be quite a milestone too. Almost a question, can intuition defeat brute force?

[towforce]

...Fable, which is the most token-hungry of the models...

I've read that Deepseek consumes 3x more tokens per task than some models, but is still cheaper overall. The leader board (link) now includes a "cost per task" column, and this is now the recommended metric.

[Bart Weststrate]

Interesting experiment!

To clarify, it was tested by a gauntlet of engines. You can find out more in the benchmark repository, that also contains the results:

https://github.com/stevemaughan/chess-engine-benchmark

— Steve

Ah so you / Claude did. I didn't read eveything there yet.

I am as we speak spsa tuning the search parameters of Selate with the fable small net. That looks very promising until now.

Thanks for your (and Claude's) hard work. And for publishing all this in detail.

[Steve Maughan]

Great works! May I study your orignal prompts? I want to try something similar on Fable too.

Here’s the original prompt in the Claude.md:

https://github.com/stevemaughan/chess-e ... /tree/main

— Steve

[Ipmanchess]

Hi Steve,

Fable 5.2 avx2 added to AI-List on R9: https://ipmanchess.yolasite.com/r9-7945 ... ngines.php

Looks like it needs +352Elo to reach Stockfish..

Ipman.