From Startup Fortune · · 1m 30s
Gemini 4 Argon Ties for Third on AI Benchmark
Google's new AI model ranks third on a benchmark it helped build but remains hidden from public tests.
In brief
Google's new AI model ranks third on a benchmark it helped build but remains hidden from public tests. Gemini 4 Argon, Google's latest AI model, ties for third on Epoch AI's benchmark but is restricted to cybersecurity partners due to limited access. Originally reported by Startup Fortune.
Gemini 4 Argon's Absence
Gemini 4 Argon hasn't appeared on Epoch AI's benchmark index because Google keeps it behind limited access.
Top Models on Index
Claude Opus 5.5 leads at 167.35, followed by GPT-6 Astra at 166.51, Claude Sonnet 5.5 at 165.2, and Claude Fable 5.1 at 164.82.
Google's Model Access
Google launched Gemini 4 Argon on September 30 but restricted it to vetted cybersecurity defenders through its Fairwind Program.
Index Methodology
The Epoch Capabilities Index uses item response theory to stitch scores from fifty benchmarks into one scale anchored by Claude 3.5 Sonnet at 130 and GPT-5 at 150.
Google Funded the Index
Google DeepMind's AGI Safety and Alignment team helped fund the ECI's methodology paper called 'A Rosetta Stone for AI Benchmarks'.
Confidence Intervals
The gap between top models sits within four-point confidence intervals, meaning a two-point edge isn't a demonstrated ordering.
Convergence of Models
Anthropic and OpenAI have converged so tightly that the ECI can't cleanly separate them, while Google's model isn't even in the running.
Google's Rollout Strategy
Google rolled out Gemini 4 Argon through Fairwind Program for cybersecurity partners, with a token limit of one million per response.
The key point
Gemini 4 Argon, Google's latest AI model, ties for third on Epoch AI's benchmark but is restricted to cybersecurity partners due to limited access.
