OpenAI's Astra/Sol 5.2 top LabBench and ScreenBench at ~20% of Fable 5.2's cost

danielmckinn0n · x · 2026-11-19

Third-party evaluator danielmckinn0n reports that OpenAI's Astra 5.2 and Sol 5.2 substantially push the Pareto frontier on GamowLabs evals, narrowly taking first on LabBench and ScreenBench. Charts show GPT-6.2 Astra leading LabBench at 50.7% and ScreenBench at 0.748, while Gemini 4 Carbon ties Claude Fable 5.2 at 61.5% on RareBench.

The most notable point is efficiency: the new models achieve this at roughly 20% of Fable 5.2's cost and token consumption, signaling a welcome shift toward brevity and speed. The author plans to compare Astra 5.2 vs Gemini 4 Hydrogen in his own workflow.

Original post →

More from Models

Models channel →