DeepSeek V4.1 Flash 1017 gains on ARC-AGI-3 while using fewer reasoning tokens
Article timeline
The articles in this story. Times are ET.
ARC PrizeOfficialARC Prize publishes DeepSeek V4.1 Flash ARC-AGI results on its leaderboardAIARC Prize has published full ARC-AGI results for DeepSeek V4.1 Flash 1017, tested with Baseten. The post links the public leaderboard, the open-source benchmarking repository for reproducing the results, and the testing policy.
- ARC PrizeOfficialDeepSeek V4.1 Flash's ARC-AGI-3 semi-private score more than doubles in new version
AIARC Prize says a new DeepSeek V4.1 Flash version, "1017," raised its ARC-AGI-3 semi-private score from 2.68% to 5.81% compared with the original model from September 10, 2026. Its public score fell from 11.05% to 8.97%, according to the post. Both runs used the provider adapter harness at max reasoning, with Baseten pricing of $0.30 per million input tokens and $1.20 per million output tokens.
- ARC PrizeOfficialDeepSeek V4.1 Flash 1017 cuts reasoning tokens on ARC-AGI games
AIARC Prize says DeepSeek V4.1 Flash 1017, run on Baseten at max reasoning, used 36% fewer reasoning tokens on its first action than the original across 25 public games. On 55 semi-private games, the model used 48.6% fewer reasoning tokens on the first action.
- ARC PrizeOfficialDeepSeek V4.1 Flash 1017 leads open-weight models on ARC-AGI-3
AIARC Prize reports DeepSeek V4.1 Flash 1017 scores 5.8% on ARC-AGI-3 at a $1.2K cost using the provider adapter harness, and 0.2% at $1.0K with the standard harness. The model leads open-weight models on ARC-AGI-3 (5.81%) and ARC-AGI-2 (78.1%), and ties the best open-weight score on ARC-AGI-1 (94.5%).
