Gemini 3 Flash uses agentic RL to rival Pro – 69% long‑context recall

Gemini 3 Flash turns out not to be a diet Pro at all. Google engineers now say it’s running on a fresh agentic RL stack, not a distilled Pro checkpoint, which helps explain why this “fast” model is punching up: SimpleBench puts Flash Preview at 61.1% (vs Gemini 2.5 Flash’s 41.2%) and Repo Bench sees ~67% real‑repo success, right in the frontier pack.

New third‑party evals sharpen the picture we started on Tuesday. Flash trades blows with Pro and GPT‑5.2 on mainstream coding and tools—78.0% vs 76.2% on SWE‑Bench Verified—while lagging on the nastiest abstraction tasks, hitting 36% on FrontierMath’s Tiers 1–3 but only 15% on Tier 4 as 5.2 pulls away. The routing story is clear: send day‑to‑day coding and broad knowledge to Flash, but keep a narrow lane on Pro or 5.2 for ARC‑AGI‑style puzzles and frontier math.

Context Arena’s MRCR runs add a practical twist. Flash Medium reaches 69.2% AUC at 128k and 45.9% at 1M tokens while burning roughly half the output cost of High, making Medium the obvious default for 128k–1M‑token agents. With Google’s new Antigravity “computer use” agent and other UI‑automation surfaces already standardizing on Flash, this agentic‑RL brain is quietly becoming Google’s real flagship for shipped work, not just benchmarks.