This is with the caveat that OpenAI uses their own harness for this:

> On ARC-AGI-3, GPT-6 Astra was run with our responses API harness , which changes two settings to better match real-world performance. The changes do not specifically target ARC-AGI-3.

Its 62 percent when using a neutral harness. https://arcprize.org/blog/astra

Their neutral harness is not very good though, if I read it right, it doesn't preserve the reasoning state between turns. No real harness discards reasoning state like that.

Yeah that must be it. OpenAI doesn't want to disclose internal reasoning, that's why thats typically encrypted_content in OpenAI codex session ledgers etc.; leveraging responses API preserves reasoning server side all the way till a final answer is made; so that's very impressive and to me the score that matters.

Yet it is an impressive number. But yeah when you see a number 99 you have doubts. Thanks for the link

Interesting both this and Sol got approximately a 37% boost with the custom harness.

This should be normalised and expected - the responses API harness allows it to use the custom compaction that is not allowed otherwise. It is entirely fair to allow OpenAI to use their own compaction algorithm..

"This should be allowed, let me explain the reason they cheated and state again that they should be allowed to cheat."

"Going forward we will capitulate and still try to keep the integrity of our benchmark in tact, but from now on every benchmark will be compromised with providers being able tweak things sufficiently to game at least a 30% bump in results."

If you mean a capability wall, the author of the benchmark says this

>We see Astra as a major breakthrough in model intelligence.

You think the author of the benchmark is also in the conspiracy

"On the current ARC-AGI-3 leaderboard, conventional frontier-model runs sit dramatically below Astra's reported 98.6% result.

But the comparison isn't straightforward.

OpenAI's own evaluation notes say Astra uses the company's Responses API harness, while comparison models can operate under different configurations."

The blog post says 99.9%. Oddly, it does better on ARC-AGI-3 than it does on version 1 or 2 of the same benchmark (though gets 95+ on all three)

I strongly suspect that is way above the human average anyway, esp. ARC 2 and 3 are really tough unless you happen to be great at those spacial puzzles or video games.

Scoring for ARC-AGI-3 is constructed so that the median(-ish) human score is 100%, so this is not a superhuman result. However, the scaling is weird, since it's built from terms that look like (AI turns taken / median human turns) ^ 2, and it weights later levels higher than early levels. So it's not at all clear that 100% is twice as good as 50%.

See the scoring docs: https://docs.arcprize.org/methodology

At this point the only valid ARC-AGI benchmark left is to make up the next series of ARC-AGI benchmark puzzles that current models presumably can't handle.

I feel like making a human-proof benchmark is pretty clear evidence that they've exceeded even the highest human capacity in most respects, for things that you can do via text generation (and to a lesser extent image generation)

Really though? I would believe something like this if a model could one shot every solution in the set. I don't pay much attention to these things and maybe this stuff is available but I would bet the session/reasoning transcript is absolutely horrendous from an intelligence standpoint.

100%, some say.-