This is fun. I finally got to follow up on a RemindMe comment. Back on March 25 of this year, six months ago, no model was getting even 1% on ARC-AGI-3. A commenter asked if we could see 75% at $2 cost. Well, the cost is still high ($26.1k), but GPT-6-Astra-Max was able to get 62.7% (no harness!)
retained text
And the commenter did technically say "a year from now", so to me that makes the 62.7% less than six months later (Astra was released earlier this month, on September 3), even more impressive. And of course with a lightweight memory adapter*, the benchmark is simply saturated. Those saturation scores are still at high cost, at least 17k, but if the trend of cheaper intelligence continues (https://epoch.ai/publications/the-plunging-price-of-thought), we should have models beating ARC-AGI-3 cheaper soon enough. "The provider adapter preserves opaque reasoning state between requests and uses compaction for longer conversations, so the model can reuse prior work rather than reconstructing it."
document fetched 2026-09-26 00:30 · canonical https://reddit.com/r/singularity/comments/1wq3qb4/this_is_fun_i_finally_got_to_follow_up_on_a