Artificial Intelligence
Gemini 4 Argon scores 97.5% on ARC-AGI-3
Google Deepmind said Gemini 4 Argon scored 97.5 per cent on ARC-AGI-3 in one harness, but its result fell sharply in another, and Epoch AI said it still lagged in coding and maths.
- Google Deepmind said Gemini 4 Argon scored 97.5 per cent on ARC-AGI-3 using the Provider Adapter Harness, at a cost of $4,800.
- ARC Prize said Google became the second AI lab after OpenAI to cross 95 per cent on the benchmark, with GPT-6 Astra at 99.9 per cent.
- In the Standard Harness, Argon scored 49.8 per cent at low reasoning and 46.3 per cent at medium reasoning, and ARC Prize did not publish a score at the highest setting because the model completed only 19 games.
- Epoch AI said Argon scored 165 on the Epoch Capabilities Index, tied Fable 5.1 and trailed Opus 5.5 and GPT-6 Astra, while ranking eighth on Math ECI and sixth on the Software Engineering Index.
Google Deepmind said Gemini 4 Argon scored 97.5 per cent on ARC-AGI-3 when it was run in the Provider Adapter Harness, at a reported cost of $4,800. ARC Prize said that put Google behind only OpenAI among the labs above 95 per cent, while GPT-6 Astra was listed at 99.9 per cent.
The result changed in the Standard Harness. There Argon fell to 49.8 per cent at low reasoning and 46.3 per cent at medium reasoning, and the testers said API problems stopped it from completing every game.
At the highest setting, the model finished only 19 of the 55 games, so ARC Prize did not publish a score for that run. The reported figures still included partial points from runs that ended early.
The article linked the setup dispute to a July 2026 comparison between GPT-5.6 Sol and Claude Opus 5. OpenAI said benchmark results also reflected the technical environment, and François Chollet later allowed general, publicly available API settings.
Epoch AI said Argon scored 165 on the Epoch Capabilities Index, equal with Fable 5.1 and behind Opus 5.5 and GPT-6 Astra. It also said the model was ahead of Gemini 3.8 Flash but placed eighth on Math ECI and sixth on the Software Engineering Index.
Named in this story
People
- François Chollet
- allowed general publicly available API settings after the dispute
Companies
- OpenAI
- was the other lab above 95 per cent and argued over the test setup
Organisations
- Google Deepmind
- reported Gemini 4 Argon's benchmark results
- ARC Prize
- published the benchmark interpretation and rankings
- Provider Adapter Harness
- was the setup that gave Argon its top score
- Standard Harness
- was the stricter setup where Argon fell
- Epoch AI
- tracked Argon's broader benchmark performance
Products and systems
- Gemini 4 Argon
- scored 97.5 per cent on ARC-AGI-3
- ARC-AGI-3
- was the benchmark used for the comparison
- GPT-6 Astra
- held the highest reported ARC-AGI-3 score
- GPT-5.6 Sol
- was part of the July 2026 benchmark dispute
- Claude Opus 5
- was the model GPT-5.6 Sol was compared with
- Responses API
- was the OpenAI interface used in the retest
- Epoch Capabilities Index
- was the composite score Argon reached 165 on
- Fable 5.1
- tied with Argon on the ECI
- Opus 5.5
- ranked just ahead of Argon on the ECI
- Gemini 3.8 Flash
- ranked behind Argon on the ECI
- Math ECI
- was the index where Argon ranked eighth
- Software Engineering Index
- was the index where Argon ranked sixth
How the source tells it
Mostly technical and comparative, with mild promotional wording around frontier status and benchmark rankings.
- competitive framing the article ranks models against rivals and lab milestones
- hype frontier language and benchmark superlatives give the piece a promotional edge