Leaked ARC-AGI-3 benchmark table shows OpenAI's unreleased GPT-5.6 Sol model scoring a record 7.8%
Posts on X are circulating a benchmark table that shows one GPT-5.6 variant far ahead on ARC-AGI-3, while a separate table raises fresh questions about its harder math performance.
A set of benchmark images circulating on X has focused attention on a post from @scaling01, which claims OpenAI's unreleased GPT-5.6 Sol scored 7.78% on ARC-AGI-3, well above Claude Opus 4.8's 1.5%. A linked benchmark post from the same account, labeled "GPT-5.6 Benchmarks", is the main source being shared for the result. A separate post from ARC Prize, reposted by François Chollet, also says GPT-5.6 Sol set a new state of the art at 7.8% and became the first verified frontier model to beat an ARC-AGI-3 game.
Combined views
564.1K
Leaked ARC-AGI-3 benchmark table shows OpenAI's unreleased GPT-5.6 Sol model scoring a record 7.8%
Posts on X are circulating a benchmark table that shows one GPT-5.6 variant far ahead on ARC-AGI-3, while a separate table raises fresh questions about its harder math performance.
Combined views
564.1K
29 posts, first seen 4h ago