GPT-6 Astra ARC-AGI-3: 62.7% vs 99.9% Explained

One headline appeared in late October: “ARC Prize rejects OpenAI’s claim that Astra scored 99.9 percent.” Startup Fortune ran it. It’s a good headline. It also doesn’t match the page it’s about.

The page is ARC Prize’s own write-up of GPT-6 Astra, published September 3 by Greg Kamradt. Its summary gives Astra 62.7% for $26K on ARC-AGI-3 Semi-Private under ARC Prize’s “Standard” setup, and 99.9% for $19K under a “Provider Adapter” one. Two numbers, one paragraph, same day as the launch. The post doesn’t reject anything.

What sits around the model

Every benchmark run has software wrapped around the model, the scaffolding that decides what the model can see, what it remembers between steps, and how long conversations get trimmed.

ARC Prize’s Standard setup is the same for every lab. The model chooses which notes to carry forward, and the interface around it stays minimal. The Provider Adapter is the opposite: it lets a vendor use its own context tools. For Astra, ARC Prize says that means keeping the model’s hidden reasoning state between requests and compacting long conversations.

ARC Prize says the two setups answer different questions. The Standard one asks how models compare under a neutral interface. The adapter asks how well a model does when it’s used the way its maker designed.

The one number I’d look at

The headline gap is 37 points, but this row is stranger. With reasoning effort set to “none”, Astra scored 35.2% in the Standard setup and 96.7% in the adapter. Turn reasoning all the way up to max in the Standard setup and you get 62.7%. So the adapter at its lowest setting beat the Standard one at its highest by about 34 points.

Astra reasoning level	Standard setup	Provider Adapter
Max	62.7% ($26,098)	98.6% ($17,332)
High	54.8% ($40,705)	99.9% ($18,817)
Low	17.5% ($38,166)	98.0% ($21,298)
None	35.2% ($49,791)	96.7% ($23,457)

Source: ARC Prize, September 3, 2026. Scores are on ARC-AGI-3 Semi-Private. Four of the six reasoning levels are shown.

The cost column runs the other way too: the cheapest Standard row in the table is the max one at $26,098, and the 99.9% adapter run cost $18,817. Across the 167 game and reasoning-level pairs both setups solved, it says the adapter runs used 49% fewer tokens and finished about 3.66 times faster.

That is a lot of performance sitting outside the model itself. Whatever you think of the 99.9%, it isn’t a fluke of one run: every adapter row sits between 96.7% and 99.9%.

Where the late-October coverage goes wrong

Startup Fortune says the 99.9% came from a “tools-enabled” run and that ARC Prize called it “not valid.” I couldn’t find either claim in ARC Prize’s September 3 post. The code sandbox in ARC Prize’s post belongs to a separate setup called PRO-LONG, and ARC Prize warns that those results combine the model and its tools. That reads to me as a separate setup from the 99.9% adapter run, though the post doesn’t draw that contrast in so many words.

ARC Prize’s own wording is different. It calls both scores state of the art, and says it will list Standard and Provider Adapter results side by side from now on, each labeled. Its leaderboard now has separate “Provider Adapter” entries for several models, Astra among them.

The fair criticism is narrower, and The Next Web made it back on September 6: the number that spread was 99.9%, and in at least one widely repeated comparison it sat against a rival score from the Standard setup. That isn’t a like-for-like comparison. Startup Fortune adds that Nvidia’s Jensen Huang posted “AGI has arrived” three days after launch, which I haven’t checked against his post.

ARC Prize isn’t claiming that. Its post says “we are not claiming that it is AGI,” and adds that ARC-AGI-3 has “deterministic, closed-ended mechanics and goals” and doesn’t represent the open-endedness of the real world.

My read

Both numbers are real, and each answers one question. If you’re comparing Astra with Claude Opus 5 or Gemini, 62.7% is the figure on equal footing. If you’re deciding whether Astra can do this kind of work when wired up properly, 99.9% is closer to the answer, assuming your own setup looks like OpenAI’s.

What I don’t like is the pattern around it. A launch that leads with the highest number, and a follow-up news cycle that leads with the most dramatic reading of the gap. Neither is lying, exactly. Both left out the one sentence that mattered, which sat in the primary source all along.

I haven’t verified the other claims in the Startup Fortune piece, such as the Prime Agent scaffold or the Nvidia engineers’ 100% result, so I’m leaving them out. For OpenAI’s other Astra headlines, see our posts on the StarCraft bot incident and the cancelled GPT-6.1 Astra.

Note on sources

I read ARC Prize’s September 3 post in full, plus its leaderboard page. The Astra results page I only saw through search excerpts, so the table above comes from the blog post. I read The Next Web’s September 6 piece and the Startup Fortune article in full. I couldn’t read OpenAI’s own launch post, so what it said about the 99.9% comes second hand from TNW.

ARC Prize's September 3 post didn't reject OpenAI's 99.9%. It printed it next to a 62.7%.