Claude Fable 5.2: 24.2 on MakoBench, tied for first

In plain terms: We ran Claude Fable 5.2 through MakoBench 4.1. It scores 24.2 and is tied for first place with Kimi K3.1 Pro, GPT-6 Astra and GPT-6.1 Sol. It beats Fable 5.1 by about five points, is faster, and refused nothing. A standard run costs $14.87, which is far more than most models near the top.

Go to: CrucibleForge on GitHub

Claude Fable 5.2 scores 24.2 ± 1.7 on MakoBench 4.1 and is tied for first place. Fable 5.1 scored 19.4 on the same suite. The full row, with every test, is on the MakoBench board.

Model	Overall	Work	Chat	Adult test	tok/s	Run cost
Kimi K3.1 Pro	24.6 ± 1.0	12.0	12.6	5.0	62.8	$4.07
Claude Fable 5.2	24.2 ± 1.7	11.5	12.7	12.0	78.9	$14.87
GPT-6 Astra	24.0 ± 1.7	11.5	12.6	1.3	41.1	$5.70
GPT-6.1 Sol	23.9 ± 1.8	11.5	12.4	1.5	111.7	$1.17
Claude Fable 5.1	19.4 ± 1.7	9.4	10.0	11.4	48.7	$10.23

Overall is Work plus Chat. The standard tests give up to 10 for each, and the bonus packs a model unlocks add the rest. The adult test is scored separately and is not part of Overall.

What changed from Fable 5.1

The overall gain over Fable 5.1 is clear of the error bars. Work rose from 9.4 to 11.5, chat from 10.0 to 12.7, and speed from 48.7 to 78.9 tokens per second. Fable 5.2 refused none of the nine refusal probes and lost no replies to a content filter. Its weak spot is agentic work: the bonus agent pack scored 0%.

Cost

A standard run of Fable 5.2 cost $14.87. GPT-6.1 Sol, tied with it, cost $1.17, and Kimi K3.1 Pro and GPT-6 Astra sit in between. Fable 5.2 is the strongest of the four on the adult test (12.0 against 5.0 or less), and the most expensive to run.

Every score, the test list and the run cost of each model are on the board. A longer comparison of the top models is coming.