Google DeepMind released a new model called Gemini 4 Argon. In a test called ARC‑AGI‑3, the model achieved a 97.5 % score. That puts it above the 95 % threshold that only a few labs have crossed. The only publicly reported model that scored higher is OpenAI's GPT‑6 Astra, which reached 99.9 %.
How the test was run
The score was obtained using a set‑up called the “Provider Adapter Harness.” This setup lets the model keep its reasoning between steps and compresses context, which is a vendor‑specific feature. When the same model was run in a more restrictive “Standard Harness,” its score fell to 49.8 % and it could not finish all 55 games.
What the numbers mean
Across the 25 public games in ARC‑AGI‑3, Argon needed 62 % fewer actions than a human baseline, matching the speed of the record in three games. On older versions of the benchmark, Argon scored 91.3 % on ARC‑AGI‑2 and 98.5 % on ARC‑AGI‑1, with costs of $0.32 and $0.16 per task respectively.
An independent index called the Epoch Capabilities Index gave Argon 165 points, tying Fable 5.1 and sitting just behind Opus 5.5 (167) and GPT‑6 Astra (166). The model leads in knowledge, agent and vision tests, but lags in math and software‑engineering tasks.
Why it matters
For developers and businesses that rely on AI for complex problem solving, Gemini 4 Argon's high ARC‑AGI‑3 score shows that Google’s models are now competitive with the very top systems. However, the lower score in the standard test and weaker math and coding performance indicate that the model may still need custom settings to reach its best results.
