GPT-6 Astra by @OpenAI achieves SOTA on ARC-AGI: - Astra scores 63%...
ARC Prize@arcprize
292 views
Sep 04, 2026
~2 min read
Advertisement
1
GPT-6 Astra by @OpenAI achieves SOTA on ARC-AGI:
- Astra scores 63% on ARC-AGI-3, 99% via a new provider adapter harness
- It surpasses human performance on 96% of ARC-AGI-3 levels
- It builds the most precise symbolic model of novel environments we've seen
Our analysis:
- Astra scores 63% on ARC-AGI-3, 99% via a new provider adapter harness
- It surpasses human performance on 96% of ARC-AGI-3 levels
- It builds the most precise symbolic model of novel environments we've seen
Our analysis:
2
Before launching ARC-AGI-3, we tested approximately 500 members of the general public to establish a human baseline for action efficiency, or simply, how quickly did people solve each environment?
This gives us an efficiency metric to compare AI to humans.
With the provider adapter harness, GPT-6 Astra surpassed the human baseline on 96% of levels.
See more of our analysis: arcprize.org/blog/astra
Full results: arcprize.org/results/openai…
This gives us an efficiency metric to compare AI to humans.
With the provider adapter harness, GPT-6 Astra surpassed the human baseline on 96% of levels.
See more of our analysis: arcprize.org/blog/astra
Full results: arcprize.org/results/openai…
3
Astra creates a dense compact symbolic world model to complete ARC-AGI-3 environments.
For example, in environment s5i5, Astra:
- Recorded the current level, hub orientation, and mechanism lengths: "L8: hub q2 (8↓). Lengths: 14=1…"
- It mapped operations to exact controls: "9−=(39,4), rotate=(49,18), 14+=(59,11)"
- And it wrote ordered plans: "extend8 to3; retract10 to2; shorten8 to1"
This on-the-fly algebraic shorthand helped Astra preserve the state that mattered and execute precise multi-step plans across turns.
For example, in environment s5i5, Astra:
- Recorded the current level, hub orientation, and mechanism lengths: "L8: hub q2 (8↓). Lengths: 14=1…"
- It mapped operations to exact controls: "9−=(39,4), rotate=(49,18), 14+=(59,11)"
- And it wrote ordered plans: "extend8 to3; retract10 to2; shorten8 to1"
This on-the-fly algebraic shorthand helped Astra preserve the state that mattered and execute precise multi-step plans across turns.
4
In addition to achieving state-of-the-art scores on ARC-AGI-3, GPT-6 Astra achieved a record 95.0% on ARC-AGI-2 at $1.12/task, and tied Fable 5's high score of 98.5% at $0.28/task.
Full results: arcprize.org/results/openai…
Full results: arcprize.org/results/openai…
5
For ARC-AGI-3 we tested GPT-6 Astra with both our Standard harness (the model decides which notes to carry forward) as well as a new provider adapter harness (which preserves opaque reasoning between requests and uses compaction).
Going forward we will test all new models on ARC-AGI-3 using both harnesses.
Read more about GPT-6 Astra's performance on ARC-AGI-3 as well as the new provider adapter harness testing changes on the ARC Prize blog:
arcprize.org/blog/astra
Going forward we will test all new models on ARC-AGI-3 using both harnesses.
Read more about GPT-6 Astra's performance on ARC-AGI-3 as well as the new provider adapter harness testing changes on the ARC Prize blog:
arcprize.org/blog/astra
6
- Leaderboard: arcprize.org/leaderboard
- Reproduce the ARC-AGI-1 and 2 results: github.com/arcprize/arc-a…
- Reproduce the ARC-AGI-3 results: github.com/arcprize/arc-a…
- Testing policy: arcprize.org/policy
- Full GPT-6 Astra results: arcprize.org/results/openai…
- Reproduce the ARC-AGI-1 and 2 results: github.com/arcprize/arc-a…
- Reproduce the ARC-AGI-3 results: github.com/arcprize/arc-a…
- Testing policy: arcprize.org/policy
- Full GPT-6 Astra results: arcprize.org/results/openai…
7
Join @fchollet, @mikeknoop, @GregKamradt, and the rest of the ARC Prize Foundation team in a forum for researchers working at the intersection of neural networks, program synthesis, and formal reasoning.
Learn more: arcprize.org/events/researc…
Attend: luma.com/z1h24dqe?tk=hp…
Apply to present: docs.google.com/forms/d/e/1FAI…
Learn more: arcprize.org/events/researc…
Attend: luma.com/z1h24dqe?tk=hp…
Apply to present: docs.google.com/forms/d/e/1FAI…
