Google DeepMind crossed 95 percent on ARC-AGI-3 with Gemini 4 Argon, scoring 97.5 percent at a cost of $4,800. Only OpenAI's GPT-6 Astra ranks higher at 99.9 percent, which makes Argon the closest Google model to the frontier in months. The result matters because ARC-AGI-3 is treated as the toughest test of general problem-solving for AI agents.
How Argon scored on ARC benchmarks
The top score came in the Provider Adapter Harness, a setup that permits vendor-specific features such as retained reasoning and context compaction. ARC Prize lists Google as the second lab to pass the 95 percent mark after OpenAI. Across 25 public games, Argon used 62 percent fewer actions than the human baseline, matching or beating the speedrun record in three games and landing within 10 percent of the best time in six others.
On earlier benchmark versions, Argon reached 91.3 percent on ARC-AGI-2 at $0.32 per task and 98.5 percent on ARC-AGI-1 at $0.16 per task. Those cost figures rest on an introductory price of $2 per million input tokens and $10 per million output tokens. The pricing detail allows buyers to compare task cost across generations rather than accuracy alone. It also shows how token pricing shapes the economics of repeated agent runs.
The same model looks different under other test rules. In the Standard Harness, which allows only simple notes between rounds, Argon scored 49.8 percent at low reasoning for $5,800. The gap follows a July 2026 dispute, when GPT-5.6 Sol beat Claude Opus 5 on ARC-AGI-3 using two API settings. OpenAI then argued that benchmarks always measure the technical environment as well, and ARC Prize co-founder Francois Chollet allowed general, publicly available API settings.
What Argon changes for AI buyers
For companies deploying agents, the Provider Adapter result signals more efficient problem-solving inside Google's stack when reasoning is preserved and context is compacted. Fewer actions per task can translate into shorter sessions, lower latency, and fewer tool calls in workflows such as research, testing, and multi-step operations. Small teams gain access to frontier-level planning without building custom orchestration, while large firms can standardize on one vendor harness across many use cases.
The limits deserve equal attention before procurement decisions. API problems stopped Argon from finishing all 55 games in the Standard Harness: 42 games at low reasoning, 46 at medium, and only 19 at the highest level, with partial points counted from aborted runs. Epoch AI places Argon eighth on the Math ECI and sixth on the Software Engineering ECI, despite a total of 165 points that ties Fable 5.1. Buyers should therefore verify math and coding performance separately and ask which harness, reasoning level, and API options support a quoted score.
The marker to watch is whether Google holds this position on the Epoch Capabilities Index, where Argon has 165 points against 167 for Opus 5.5 and 166 for GPT-6 Astra. Argon already moved well ahead of Gemini 3.8 Flash at 157 points, the best standing since Gemini 3.1 Pro in February. A stable top-three place across knowledge, agent, and vision benchmarks would confirm that the gap to the frontier has closed for business workloads.
