Google has unveiled Gemini 4 Argon — its new frontier AI model, with a 97.5% score on the ARC-AGI-3 benchmark, placing it second only to OpenAI’s GPT-6 Astra (99.9%) among publicly reported models. The result, confirmed by ARC Prize, marks Google’s strongest showing on that benchmark since Gemini 3.1 Pro in February. And the first time a Google model has crossed 95% on ARC-AGI-3 since the benchmark’s July 2026 revision made it harder. But the number comes with caveats: it was achieved using the “Provider Adapter Harness,” a test setup that leverages Google-specific API features like retained reasoning and context compaction. Under the stricter “Standard Harness”, which resets internal state between actions and uses only basic API calls. Argon drops to 49.8% at low reasoning, and failed to complete all 55 games at higher settings. At maximum reasoning, it finished just 19 games. No official score was published for that tier. Epoch AI’s broader Epoch Capabilities Index (ECI) gives Argon 165 points, tying Fable 5.1, trailing Opus 5.5 (167), and edging ahead of GPT-6 Astra (166) by one point. It beats Gemini 3.8 Flash (157), but lags behind rivals in math (8th) and software engineering (6th) sub-scores. Pricing is not public. ARC Prize reports the Provider Adapter Harness run cost $4,800. The Standard Harness run cost $5,800. But those are research evaluation costs, not commercial API pricing. No launch date, availability window, or integration timeline into Gemini apps, Android, or Google Cloud has been announced. What’s clear is this: Argon isn’t shipping yet. It’s a lab milestone, one that exposes how much benchmark results now depend on how you test, not just what you test.