AI News
All stories
Reported

GPT-6.2 Astra and Sol top MathArena, while Astra places sixth in Agent Arena

Astra and Sol took the top two spots on MathArena, and Astra solved 42 of 53 ArXivLean problems, according to the third-party results cited.

Models · · 1 min read

Astra and Sol took the top two places on MathArena, and Astra solved 42 of 53 ArXivLean problems, according to the third-party results cited.

Outside formal math, the picture changes. On Agent Arena, Astra debuted sixth and Sol placed eleventh at a $0.90 median task cost. On the separate Code Arena WebDev board, Astra placed third at a 42% lower median cost than its predecessor. OpenAI also opened a public beta for GPT-6.2 Responses API jobs that can wait. Two-hour and eight-hour targets for the first response carry discounts of 30% and 40% per token, respectively. The cheaper pricing is therefore tied to accepting a delay.

OpenAI's Tibo Sottiaux says early reviewers used builds from before the final few days of changes. Artificial Analysis, in a behavioral evaluation, found that Astra's responses became shorter and that it was less likely to admit uncertainty. Those findings concern how the model behaves rather than its scores. Jeremy Howard said he was reverting Solveit's Astra and Sol options to their pre-6.2 versions.

What changed

OpenAI opened a public beta of GPT-6.2 Responses API jobs that can wait, with 30% and 40% per-token discounts for two-hour and eight-hour first-response targets.

Type: New capability · Applies to: GPT-6.2 Responses API

Why it matters

Benchmark standings vary sharply by task. The models lead on formal math but trail on agent cost and behavioral reliability, so model choice should follow the specific workload.

Original source

We have no link to the original source for this story yet.

Where we saw it

1 source

Latent Space

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works