Google’s Gemini Gets a Reasoning Upgrade — and This Time It Looks Bigger Than a Model Refresh
Google’s latest Gemini push appears to be a two-step upgrade, not a single launch. On February 12, Bloomberg reported that Google had updated Gemini Deep Think to improve performance in math and science research. A week later, TechCrunch reported that Google released Gemini 3.1 Pro in preview, bringing stronger reasoning into its broader Pro model lineup. (Bloomberg.com)
What makes the move notable is its practical target. Bloomberg reported that Google developed the new Deep Think in close partnership with researchers and is positioning it to help move work from theoretical reasoning into practical applications. VentureBeat reported that Gemini 3.1 Pro is aimed at science, research, and engineering workflows that require deeper planning and synthesis. Reuters, meanwhile, reported earlier this month that Gemini had already grown to more than 750 million monthly active users and 8 million paying enterprise licenses by the end of Google’s December quarter, giving the upgrade a very large installed base from day one. (Bloomberg.com)
The headline metric behind the release is ARC-AGI-2, a benchmark built to test novel abstract reasoning. ARC Prize’s leaderboard lists Gemini 3.1 Pro Preview at 77.1% and Gemini 3 Deep Think at 84.6%, while also noting that preview results can be unofficial or based on incomplete testing. VentureBeat reported that the 77.1% figure is more than double Gemini 3 Pro’s earlier 31.1% score, and Artificial Analysis said Gemini 3.1 Pro Preview moved to the top of its Intelligence Index, ahead of Claude Opus 4.6. (ARC Prize)
The case for the upgrade does not rest on a single benchmark. TechCrunch reported that Google also highlighted gains on Humanity’s Last Exam, a test whose own site describes it as a 2,500-question, multi-modal benchmark designed to remain difficult as older evaluations become saturated. Artificial Analysis said Gemini 3.1 Pro Preview leads six of the ten evaluations in its index, including scientific reasoning, coding, and a measure of hallucination reduction. (TechCrunch)
Access is widening, but not evenly. Bloomberg reported that the upgraded Deep Think is available in the Gemini app for Google AI Ultra subscribers, with API access being opened selectively to researchers, engineers, and enterprises. VentureBeat reported that Gemini 3.1 Pro is rolling out in preview across the Gemini app, NotebookLM, Vertex AI, and the Gemini API. In other words, Google is not keeping improved reasoning as a lab demo; it is pushing it into both consumer products and enterprise workflows. (Bloomberg.com)
The larger story is that Google is trying to make reasoning less of a premium side feature and more of a baseline capability across Gemini. Independent trackers say the new model is leading several major evaluations, and Reuters says Gemini already operates at massive user scale. That does not prove benchmark wins will translate cleanly into everyday usefulness, but it does mean this release matters for more than leaderboard optics. (Artificial Analysis)
