Google's Gemini 4 Argon is closer to the AI frontier than any Google model in months

Google DeepMind's new frontier model, Gemini 4 Argon, has become the second major AI lab to pass the 95 percent mark on the ARC-AGI-3 benchmark, according to results reported by ARC Prize and analyzed by the research institute Epoch AI. Argon scored 97.5 percent on the test, placing it behind only OpenAI's GPT-6 Astra at 99.9 percent. The achievement makes Google the second lab after OpenAI to cross that threshold on what is widely regarded as the toughest general problem-solving benchmark currently in use.
The headline score came at a cost of $4,800 and depended on a specific testing configuration called the "Provider Adapter Harness," which draws on vendor-specific API features such as retained reasoning and context compaction. Retained reasoning means the model keeps its internal chain of thought available across steps rather than discarding it after each action; compaction intelligently summarizes and trims context to keep it within token limits. Those features can meaningfully change how a model performs on multi-step tasks, because they let it build on earlier conclusions instead of starting fresh each round.
The gap between test setups is large. In the "Standard Harness," which permits only simple notes between rounds, Argon's score falls to 49.8 percent at low reasoning effort, at a higher cost of $5,800. That is roughly half the headline figure, a spread that illustrates how sensitive benchmark results can be to infrastructure choices.
That sensitivity is not new. The source traces it to a July 2026 dispute in which OpenAI's GPT-5.6 Sol was reported to beat Claude Opus 5 on ARC-AGI-3 using two API settings. Under the standard environment, GPT-5.6 Sol managed only 7.8 percent, because that setup discards internal reasoning after each action and used an older interface lacking newer Claude API features. That meant the two models were tested under different conditions. When GPT-5.6 was tested with OpenAI's own Responses API, including retained reasoning and smart context compaction, it jumped to 38.3 percent and beat Opus 5 at 30.2 percent. OpenAI's argument was that benchmarks always measure the technical environment alongside the model. ARC Prize co-founder Francois Chollet subsequently allowed general, publicly available API settings in the evaluation.
Even in the favorable harness, API problems prevented Argon from completing all 55 games in the Standard Harness, according to the testers. At low reasoning effort the system finished 42 games and scored 49.8 percent; at medium effort it finished 46 games and scored 46.3 percent. At the highest reasoning level it completed only 19 games, so ARC Prize did not publish a score. The reported results still cover all 55 games by including partial points from aborted runs.
On gameplay efficiency, Argon performed strongly. Across the 25 public games of ARC-AGI-3, it needed 62 percent fewer actions than the human baseline, matching GPT-6 Astra on that measure. It matched or beat the speedrun record in three games and came within 10 percent of the best time in six others. On older versions of the benchmark, Argon scored 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, based on an introductory price of $2 per million input tokens and $10 per million output tokens. That pricing is notably low for a frontier model.
Epoch AI tracked the model on its own scale. Argon scored 165 points on the Epoch Capabilities Index (ECI), tying Fable 5.1 and landing just behind Opus 5.5 at 167 and GPT-6 Astra at 166. The ECI aggregates more than 50 benchmarks into a single capability scale, similar in spirit to how Artificial Analysis builds its main score. Argon jumps well ahead of Gemini 3.8 Flash at 157 points and brings Google closer to the top of the field than at any point since Gemini 3.1 Pro in February 2026. Artificial Analysis and Vals.ai reportedly confirmed the trend.
Domain performance is uneven, however. According to Epoch AI, Argon leads several knowledge, agent and vision benchmarks, but early estimates place it eighth on the Math ECI and sixth on the Software Engineering Index (SWE ECI). For developers choosing a model for coding agents or quantitative work, that means Argon may not be the strongest option despite its overall frontier position. The ECI spread between 165 and 167 points is small, so small benchmark differences at this level should be read with caution; a few points rarely translate into a decisive practical advantage.
The broader lesson is that benchmark scores increasingly reflect the evaluation harness as much as the model. Vendor-specific features like retained reasoning and context compaction are not universally available across providers, so a score achieved with them may not be reproducible through a generic API. For teams comparing models, this suggests reproducing results under your own stack rather than relying on headline numbers alone. ARC-AGI-3 was designed to resist memorization by requiring novel problem-solving, but how a model is queried still shapes what it can do.
Why it matters: Google has closed much of the gap to the frontier, giving product teams a credible alternative to OpenAI and Anthropic for agentic and reasoning workloads, especially at the stated introductory token prices, though the pricing is described as introductory and may not last. Developers evaluating Argon should note that its strongest scores come from a provider-specific harness, and that it ranks lower on math and software engineering benchmarks, so coding-heavy or quantitative applications may still favor rival models. This is inference, but teams that benchmark models against their own tooling rather than published leaderboards are likely to make better decisions here.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Google Deepmind's Gemini 4 Argon scored 97.5 percent on the ARC-AGI-3 benchmark, making Google the second lab after OpenAI to cross the 95 percent mark. But the results swing wildly depending on the test setup: in the standard harness, the score drops to just 49.8 percent. Argon also trails rivals in math and software engineering. The article Google's Gemini 4 Argon is closer to the AI frontier than any Google model in months appeared first on The Decoder .