What score will Gemini 3.5 Pro achieve on Epoch AI's primary aggregate capabilities index (the Epoch Capabilities Index, or Epoch's current headline aggregate), using its maximum publicly available configuration, within 60 days of public release?

Metric Context & Current Frontier Epoch AI’s Epoch Capabilities Index (ECI) is a comprehensive aggregate of ~39–41 benchmarks, calibrated such that Claude 3.5 Sonnet = 130 and GPT-5 = 150 epoch.aiepoch.aiepoch.ai. The index is currently heavily compressed at the upper end: the state-of-the-art (SOTA) is anchored by Claude Fable 5 at 161 and GPT-5.5 Pro at 159 epoch.ai. Crucially, Google’s lightweight Gemini 3.5 Flash already scores an impressive 155 epoch.ai, while the previous-generation Gemini 3.1 Pro sits at 156 epoch.ai.

Capability Drivers for Gemini 3.5 Pro Gemini 3.5 Pro, integrated with its maximum "Deep Think" configuration, should comfortably match or marginally exceed the current SOTA. A standard "Pro" tier step-up historically yields a measurable ECI gain over its Flash counterpart. Furthermore, Epoch’s evaluation methodology is highly favorable for this specific release because it selects the highest benchmark scores across available thinking settings epoch.ai. This ensures the index will fully capture the test-time compute advantages generated by Deep Think.

We see substantial room for ECI point gains on specific, highly weighted component benchmarks. Currently, Gemini 3.5 Flash lags behind the top frontier on complex reasoning tasks; for instance, it scores 26.8% on FrontierMath Tier 4 v2, compared to GPT-5.5 Pro’s 78.0% epoch.ai. A flagship release running maximum-effort Deep Think is explicitly designed to close these math and science reasoning gaps, which will translate to a robust aggregate boost on the ECI.

Risks and Uncertainty Our median expectation centers at 161.2, assuming Google aims to meet or slightly push past the current 161 SOTA limit upon its rollout.

• Downside (p10/p25 = 156.8–158.8): Top-end index compression structurally limits absolute point gains. Alternatively, the public Pro release could be more optimized for latency and product integration than raw capability, or Epoch may lack sufficient high-effort Deep Think runs within the 60-day window, resulting in a score that merely reflects a modest bump over the baseline 3.5 Flash.
• Upside (p75/p90 = 163.5–165.8): The integration of heavy test-time compute acts as a genuine generational leap, closing all major reasoning gaps and pushing the model significantly beyond the currently packed 159–161 frontier, setting a definitive new high-water mark on the ECI scale.