جیمنی 3.1 پرو

Google's frontier reasoning model, built to refine the performance and reliability of the Gemini 3 Pro series with better thinking, improved token efficiency and a more grounded, factually consistent experience. It offers advanced problem-solving and strong agentic and coding capability, and is strongest on agentic performance, advanced coding, long-context understanding and algorithmic development. It reasons on every request across a million-token multimodal context.

Pro is the tier Google reserves for problems where a fast answer is the wrong answer, and 3.1 Pro is the core intelligence underneath the Deep Think update the lab had released the week before. Google frames it not as a new size class but as a smarter baseline: the same Gemini 3 Pro shape, with the reasoning rebuilt underneath it.

The headline number is ARC-AGI-2, which scores a model on logic patterns it has never seen. Google reports a verified 77.1%, against 31.1% for Gemini 3 Pro - more than double, on the one evaluation designed to resist memorisation. The rest of the lab's table is broad rather than pointed: 94.3% on GPQA Diamond, 80.6% on SWE-Bench Verified and 54.2% on SWE-Bench Pro at a single attempt, 68.5% on agentic terminal coding, 59% on SciCode, and a competitive programming Elo of 2887 against 2439 for its predecessor. On agentic breadth it reports 69.2% on multi-step tool workflows against 54.1%, 85.9% on agentic search against 59.2%, and 33.5% on long-horizon professional tasks against 18.4%.

What Google chose to demonstrate says as much as the scores. The examples are all things that are code underneath: animated vector graphics generated straight from a prompt and kept small because they are code rather than pixels, a live dashboard wired to a public telemetry stream, an interactive simulation driven by hand tracking, and a designed website reasoned out of a novel's tone. The pattern is a model asked to build something that has to run, not to describe it.

Read the preview label as a real one. Google shipped 3.1 Pro in preview on purpose, to validate the update and to keep working on ambitious agentic workflows before general availability, and it is the priciest Gemini text model here by some distance. Its own table also shows the limits: multimodal understanding sits fractionally below Gemini 3 Pro, expert-task Elo trails the strongest competing models Google lists, and at the full million-token length the pointwise long-context score is unchanged from the previous generation. A million tokens of window is not a million tokens of comprehension, and Google's own numbers say so.