Claude Fable 5.2 debuts in first place
In GameBench 2, AI models write programs for eight strategy games, and the programs then play each other. Anthropic released Claude Fable 5.2 on 9 November, and we began generating its programs the same day. In our release of 11 November it is first in the Overall ranking, which averages a model's three reasoning settings, with 95.0. GPT-6 Astra follows with 88.8, Claude Opus 5.5 with 87.6 and Claude Fable 5.1 with 87.1. At XHigh, Claude Fable 5.2's programs score 99.1, the highest of the 258 ranked model-and-setting rows. They won 4,452 of their 4,484 matches.
Setting by setting
We score up to three reasoning settings for each model. For these four models they are Low, Medium and XHigh. Each cell below gives the score, the rank among all 258 ranked rows (for the Overall line, among the 73 models with an Overall rank) and the cost per program.
Scroll sideways to see every column.
| Setting | Claude Fable 5.2 | Claude Opus 5.5 | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| Low | 91.8 (7), $0.47 | 76.6 (32), $0.12 | 84.7 (18), $0.79 | 83.4 (20), $0.24 |
| Medium | 94.0 (5), $0.65 | 91.4 (9), $0.33 | 82.9 (22), $1.15 | 88.6 (12), $0.41 |
| XHigh | 99.1 (1), $4.58 | 94.8 (2), $1.99 | 94.0 (4), $5.15 | 94.3 (3), $1.30 |
| Overall | 95.0 (1), $1.90 | 87.6 (3), $0.81 | 87.1 (4), $2.32 | 88.8 (2), $0.65 |
Claude Fable 5.2 leads each of the other three models at every setting. The XHigh gap is 4.3 points to Claude Opus 5.5, 4.8 to GPT-6 Astra and 5.1 to Claude Fable 5.1. Its Medium programs, at 94.0, come within a point of all three of those XHigh rows, and its Low programs score 0.4 points above the best Medium row of the other three, Claude Opus 5.5's.
Game by game, Claude Fable 5.2's XHigh programs have the highest score in 7 of the eight games. Their game scores run from 97.6 in Go to 99.7 in Xiangqi. They won all 544 of the Xiangqi matches and all 554 in International Draughts.
Twenty-four working programs
Claude Fable 5.2 has a working program for all 24 of its game-and-setting pairs, and none of them needed a repair request. For 23 of them the first request was enough. At XHigh, one program needed more than one request, because the model used up its 128,000 output tokens before it had written any code.
Cost per program
A cost per program is the mean cost of generating one program, with repairs and failed attempts included. For these four models the costs are estimates from public list prices, not provider bills. Anthropic lists Claude Fable 5.2 at $8 per million input tokens and $40 per million output tokens, against $10 and $50 for Claude Fable 5.1.
The release puts an XHigh program from Claude Fable 5.2 at about $4.58, 2.3 times the cost of Claude Opus 5.5's and 3.5 times GPT-6 Astra's. That figure averages the repeated generation runs of the program that needed more than one request. With every request counted in full at list price, an XHigh program costs about $5.67, more than the $5.15 of Claude Fable 5.1. Within Claude Fable 5.2, XHigh costs 7.1 times as much as Medium and scores 5.1 points more. At Medium a program costs $0.65 and scores about what the other three models score at XHigh, where their programs cost $1.30 to $5.15.
Reading the score
A program's 0–100 score is its expected share of the points against the reference field, a fixed set of programs that every program is scored against. At 99.1, Claude Fable 5.2's XHigh row is close to the top of that scale. The number describes its strength against today's field, and a stronger model would have less than a point to gain there.
The release covers 105 models, 2,237 programs and 627,550 rated matches. The leaderboard always shows the newest release, Claude Fable 5.2's model page has its game-by-game results, and the methodology explains the scoring in full.