Google releases Gemini 2.0 Flash Thinking, its first public reasoning model

Google released Gemini 2.0 Flash Thinking Experimental, its first publicly available model trained to reason through a problem step by step before answering, a week after shipping Gemini 2.0 Flash and roughly two weeks after OpenAI made the full o1 reasoning model widely available. Built on the Flash model’s speed, it displayed its intermediate reasoning to users, who could expand a “thoughts” panel to see the steps behind an answer before it was given — a more visible presentation of chain-of-thought than o1 offered, where OpenAI kept the underlying reasoning tokens hidden from users by design.

Google reported that the model debuted at first place across every category of the Chatbot Arena leaderboard, a crowdsourced ranking based on pairwise human preference votes, including the overall ranking, a style-controlled variant intended to reduce the influence of answer formatting on votes, mathematics, creative writing and hard prompts. The company did not publish comparisons against o1 on task-specific benchmarks such as competition mathematics or coding, where OpenAI’s own reported figures had been strongest.

Google described the release explicitly as an early, experimental model rather than a finished product, available through Google AI Studio and Vertex AI without a defined roadmap for when — or in what form — a reasoning-optimised model would reach the consumer Gemini app. The release established reasoning as a second front, alongside raw model capability, on which Google and OpenAI were now competing directly and on a similar timetable, a contest that continued through 2025 as both companies iterated toward more capable reasoning models.