Key points
- Epoch AI ranks Gemini 4 Argon near the top
- Where Epoch says Argon still lags
- How ARC Prize tested Argon on ARC-AGI-3
Alphabet's (GOOGL) Gemini 4 Argon scored 165 on the Epoch Capabilities Index, the closest a Google model has come to the top of that ranking since Gemini 3.1 Pro in February. Argon ties Anthropic's Claude Fable 5.1 and sits behind Claude Opus 5.5 at 167 and OpenAI's GPT-6 Astra at 166, according to Epoch AI.
ARC Prize, which runs the ARC-AGI reasoning benchmarks, separately scored Argon at 97.5% on its newest test, ARC-AGI-3, using what ARC Prize calls its provider adapter harness. Both groups posted their results on Oct. 15, two days after Google opened Argon to developers and paid subscribers.
Epoch says Argon lags in math and software engineering
Epoch AI combines results from many benchmarks into its index. Argon's 165 is up from 157 for Gemini 3.8 Flash. "It tops some knowledge, agentic, and image benchmarks, while lagging behind in math and software engineering, ranking 8th on Math ECI and 6th on SWE ECI in early estimates," Epoch wrote.
On the Artificial Analysis Intelligence Index, Argon scored 53 as of Friday, tied with GPT-6 Astra and Claude Fable 5.1 and behind Claude Opus 5.5 at 58 and Claude Sonnet 5.5 at 56. Argon also scored 53 on Oct. 2, when Claude Opus 5.5 scored 58 and Claude Sonnet 5.5 scored 56.
Different test setups, with some runs incomplete
ARC-AGI-3 is a set of 55 interactive games. ARC Prize tested Argon with two harnesses, the software that manages how a model interacts with the games and retains information between turns. The provider adapter harness "preserves opaque reasoning and enables automatic compaction," while the standard harness "enables the model to carry forward notes between turns," ARC Prize said.
With the provider adapter harness, Argon scored 97.5% at low reasoning. With the standard harness, it scored 49.8% at low reasoning and 46.3% at medium. API problems kept ARC Prize from finishing those runs. Argon completed runs on 42 of the 55 games at low reasoning and 46 at medium. At high reasoning, runs finished on only 19 games, so ARC Prize didn't publish a score.
GPT-6 Astra holds the top ARC-AGI-3 score on the ARC Prize leaderboard, 99.9% with the provider adapter harness, at a reported cost of $18,800. Argon's 97.5% cost $4,800. ARC Prize priced Argon at Google's introductory rates of $2 per million input tokens and $10 per million output tokens, which are half its standard rates. The leaderboard lists no ARC-AGI-3 score for Claude Opus 5.5 or Claude Fable 5.1.
On the older ARC-AGI-2 test, Argon scored 91.3% at high reasoning. The best scores for GPT-6 Astra, GPT-6.1 Sol, and Claude Opus 5.5 were 95.0%, 94.2%, and 93.3%.
Alphabet shares were down 1.7% at $343.26 as of 12:30 p.m. ET Friday.
