Google DeepMind's Gemini 4 Argon scores 97.5% on ARC-AGI-3 under vendor-specific harness, 49.8% under standard harness

ARC Prize reported that Google's Gemini 4 Argon scored 97.5% on ARC-AGI-3 at a cost of $4,800 when run with a vendor-specific Provider Adapter Harness. Under the Standard Harness it scored 49.8% at low reasoning ($5,800), and API problems stopped it completing all 55 games, so there is no published score at the highest reasoning level. It also scored 91.3% on ARC-AGI-2 and 98.5% on ARC-AGI-1. Epoch AI rates it 165 on its Capabilities Index, just behind Opus 5.5 (167) and GPT-6 Astra (166), but eighth on its Math index and sixth on its software engineering index. Introductory pricing is $2 per million input tokens and $10 per million output tokens.

Google now looks competitive with the top two labs on general capability, which gives developers and enterprises a credible third option and may put pricing pressure on rivals. The $2/$10 token pricing is the most directly usable fact, though it is introductory and may change. Teams choosing a model for maths or coding work should not read this as a reason to switch, because Argon trails on both. The gap between the two harness results means the headline score tells you little about how it will perform in your own setup.

The 97.5% figure is the least informative number here. It comes from a vendor-specific harness, while the standard harness gave 49.8%, so "second lab past 95%" says more about tooling than about the model. The "closest to the frontier in months" framing rests on an aggregate index where Argon is still behind two rivals, and the article gives no independent replication or primary source links.

1. 01Whether ARC Prize publishes a complete Standard Harness score at the highest reasoning level once the API problems are fixed, and how it compares with 49.8%.
2. 02Independent results on real coding and maths tasks (SWE-style benchmarks, Artificial Analysis rankings) to see if the sixth and eighth place Epoch positions hold up.
3. 03Whether the $2/$10 introductory pricing is kept after the launch period, and whether OpenAI or Anthropic cut prices in response.