"GPT-5.6 Sol sets a new SOTA on ARC-AGI-3: 7.8% Sol is the first verified frontier model to ever beat an ARC-AGI-3 game It is the best model at orienting in a situation it's never encountered" — ARC Prize
SUMMARYARC Prize said GPT-5.6 Sol became the first verified frontier model to beat an ARC-AGI-3 game, solving FT09 in 152 actions versus a 208-action human baseline. The model showed strong scene comprehension and made progress on ARC-AGI-2 and ARC-AGI-1 as well, with Sol, Terra, and Luna pushing the performance-cost frontier for reasoning tasks.
GPT-5.6 is the first model to show material progress on ARC-AGI-3
See full results: https:// arcprize.org/results/openai -gpt-5-6 …
Sol beats ARC-AGI-3 game (FT09) in 152 actions compared to 208 human baseline
See the replay: https:// arcprize.org/replay/e726990 3-8865-4616-a0ee-ab6f99e328d7 … Play FT09 yourself: https:// arcprize.org/tasks/ft09 Sol’s distinguishing capability is scene comprehension
It almost always figures out what the core game mechanics actually are (unlike other models)
Sol discovers a complicated mechanic in LP85 where game pieces must be held or parked until they are needed:
“The horizontal When analyzed against Opus 4.8, GPT-5.6 Sol was able to quickly grasp the connection mechanic in CN04 Though Sol was able to beat FT09, on BP35 it wasn’t able to complete a level
The observed failure isn't perception. It reads the board correctly (same as in the games it wins)
What breaks down is Sol's reasoning on top: as the required chain of inference gets deeper, the model Sol, Terra and Luna all push the pareto frontier towards more efficient performance
ARC-AGI-2 * Sol: 92%, $1.44/task * Terra: 83.9%, $1.09/task * Luna: 59.5%, $0.67/task
ARC-AGI-1 * Sol: 96.5%, $0.54/task * Terra: 96.5%, $0.55/task * Luna: 88%, $0.32/task — ARC Prize