Gemini 3 Deep Think: Science Update, Access & Benchmarks

A research mode reached more working scientists

On February 12, Google updated Gemini 3 Deep Think for AI Ultra subscribers in the Gemini app. Selected researchers and organizations could also seek early API access. The announcement did not open a generally available API to every developer. [1]

The Google DeepMind team had introduced Gemini 3 Deep Think in November 2025 as an enhanced reasoning mode, initially provided to safety testers. February was an upgrade to that product line, not the invention of a separate pay-per-answer service. [2]

Google's earlier July 2025 research version illustrated one technique behind Deep Think: explore several possible solutions in parallel and combine them before answering. That is a concrete use for additional reasoning effort, rather than simply writing a longer response. The February announcement did not disclose an equivalent technical breakdown of its updated mode. [3][1]

From a sketch to something that can be checked

Google described a sketch-to-3D-printing workflow: the model interprets a drawing, represents the shape and produces a printable file. It also reported that an early tester used the system to identify a logical flaw in a mathematics paper. [1]

Atlas interpretation: The interesting output is an artifact someone can inspect. A generated part needs dimensions and tolerances checked; a proposed correction to a proof needs the argument reworked. These examples suggest ways to accelerate an iteration, but the announcement supplied no engineering success rate. [1]

Competition results set a narrower standard

The earlier July 2025 mathematical olympiad result used an advanced, specialized Deep Think version. Google reported that it solved five of six problems, with grading by official IMO coordinators. That result concerned a particular system and contest setup, not every later Deep Think request. [3]

For the February update, Google reported 48.4% on Humanity's Last Exam without tools. The tool condition matters: a model answering unaided is being tested differently from a research workflow that can retrieve sources or run calculations. [1]

Atlas interpretation: Olympiad problems have a defined statement and a solution that can be graded. Research adds another difficulty: deciding whether the question is well posed and whether the available evidence supports an answer at all. Competition performance is useful evidence about problem solving, but it cannot supply missing observations. [3]