One fixed score on a reasoning puzzle test went from $200 a puzzle to about a penny
Passing 75% on a set of visual puzzles cost $200 a puzzle in December 2024 and $0.012 by July 2026. The best models still cost far more.
Cost per task to score 75%+ on ARC-AGI-1
Cheapest entry on ARC Prize's leaderboard scoring at least 75% on the ARC-AGI-1 semi-private set, USD per task. Log scale.
What this measures: How much it costs an AI to solve one ARC-AGI-1 visual logic puzzle while scoring 75% or more.
Dates are ARC Prize's model release dates. Costs depend on ARC Prize's pricing assumptions, which it has revised (o3-preview went from $26 to $200 per task). The tested o3-preview was trained on part of the public training set and differs from the released o3.
Show data table and sources
| Date | Label | Value |
|---|---|---|
| 20 Dec 2024 | o3-preview (low) | $200 |
| 7 Aug 2025 | Grok 4 refinement system | $8.42 |
| 18 Nov 2025 | Gemini 3 Pro | $0.49 |
| 17 Dec 2025 | Gemini 3 Flash Preview (high) | $0.17 |
| 9 Jul 2026 | GPT-5.6 Terra (medium) | $0.13 |
| 21 Jul 2026 | Gemini 3.6 Flash (low) | $0.12 |
| 30 Jul 2026 | GPT-5.6 Luna (high) | $0.028 |
| 31 Jul 2026 | DeepSeek V4 Flash (low) | $0.012 |
- ARC-AGI Leaderboard (ARC Prize)
- ARC Prize leaderboard data (evaluations.json) (ARC Prize)
- OpenAI o3 Breakthrough High Score on ARC-AGI-Pub (ARC Prize)
ARC-AGI-1 is a set of visual reasoning puzzles built to be easy for people and hard for AI. In December 2024 ARC Prize reported OpenAI’s o3-preview as the first to pass 75%, scoring 75.7%. The run cost $26 per puzzle at the time, which ARC Prize now lists as $200 after repricing it.
What got cheap is that fixed score, not AI in general. The cheapest model on the leaderboard reaching 75% or better fell to $8.42 (August 2025), then $0.49 (Gemini 3 Pro, November 2025), $0.17 (Gemini 3 Flash, December 2025) and $0.0118 (DeepSeek V4 Flash, scoring 84.0%, July 2026).
The best models are still expensive, because the frontier moved to harder puzzles. The top score on ARC-AGI-2 is 95.0%, by GPT-6 Astra at maximum effort, and that costs $1.12 per puzzle.
What it means
An ability worth paying $200 a go for in 2024 now costs about a penny, so anything built on that level of reasoning is cheap enough to run at scale. This is about a fixed level of ability getting cheaper, not the best AI getting cheaper: the newest models on harder tests still cost dollars per puzzle. If you want today’s best, the price has not fallen.
Caveats
- The costs rest on ARC Prize’s own pricing assumptions, which it has revised, so the $200 figure is a later estimate rather than what was paid in 2024.
- The o3-preview that was tested had seen part of the public training set and differs from the o3 that was released.
- ARC-AGI-1 and ARC-AGI-2 are close to saturation, meaning top scores leave little room to improve, so they will stop telling us much.