Gemini 2.5 Deep Think
Last reviewed
Jun 3, 2026
Sources
6 citations
Review status
Source-backed
Revision
v1 · 1,282 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
6 citations
Review status
Source-backed
Revision
v1 · 1,282 words
Add missing citations, update stale details, or suggest a clearer explanation.
Gemini 2.5 Deep Think is an enhanced reasoning mode for Google DeepMind's Gemini 2.5 Pro model that uses "parallel thinking" to explore multiple solution paths at once before settling on an answer. It was previewed at Google I/O in May 2025 and released to Google AI Ultra subscribers in August 2025. The mode is closely tied to an advanced internal version of Gemini with Deep Think that earned a gold-medal score at the 2025 International Mathematical Olympiad, though Google DeepMind has been careful to distinguish the consumer product from that research system.
Deep Think was unveiled on May 20, 2025, during the Google I/O developer conference, where Google described it as an experimental reasoning capability for Gemini 2.5 Pro built on "new research techniques enabling the model to consider multiple hypotheses before responding." [1] At launch it was not generally available. Google opened it first to trusted testers through the Gemini API to collect feedback before a wider rollout. [1]
The feature sits in the broader family of "thinking" models, which spend additional compute reasoning through a problem before producing output. Where a standard Gemini 2.5 response is relatively fast, Deep Think trades latency and cost for accuracy on harder problems in mathematics, coding, and scientific analysis. Google positioned it as its most capable reasoning configuration short of the dedicated research systems used for competition mathematics.
Deep Think reached general availability for paying users on August 1, 2025, when it began rolling out to subscribers of the Google AI Ultra plan inside the Gemini app on web, Android, and iOS. [2][3]
The defining technique behind Deep Think is parallel thinking. Rather than following a single chain of reasoning, the model generates many candidate ideas simultaneously, weighs them against one another, and can revise or combine them over time before choosing a final answer. [2] Google DeepMind framed it as a system that "doesn't just answer, it brainstorms," combining parallel exploration with reinforcement learning techniques. [4]
TechCrunch described the released product as Google's "first publicly available multi-agent model," noting that it spawns multiple agents to tackle a question in parallel. [3] That approach consumes substantially more computational resources than a single-agent response but tends to produce better results on difficult tasks. The trade-off is reflected in the strict usage caps Google placed on the consumer version.
In normal use Deep Think can automatically call tools, including code execution and Google Search, and it generates considerably longer responses than the standard model. [2][3] Google highlighted its strength on iterative software development, web design, scientific discovery, and the formulation of mathematical conjectures.
The most prominent result associated with Deep Think came at the 2025 International Mathematical Olympiad. On July 21, 2025, Google DeepMind announced that an advanced version of Gemini with Deep Think had achieved a gold-medal standard, solving five of the six problems for 35 points out of a possible 42. [5] The IMO itself graded and certified the solutions using the same criteria applied to student contestants. IMO President Gregor Dolinar confirmed the outcome, stating that "Google DeepMind has reached the much-desired milestone, earning 35 out of a possible 42 points, a gold medal score." [5]
A crucial detail is that this was an advanced research system, not the product later shipped to subscribers. The model operated end-to-end in natural language, reading the official problem statements and writing rigorous proofs directly within the 4.5-hour competition time limit. [5] That marked a step beyond Google DeepMind's 2024 effort, when the combined AlphaProof and AlphaGeometry 2 systems reached the silver-medal standard with four problems and 28 points, but required human experts to first translate the problems into formal languages such as Lean. [5] The 2025 system removed that translation step entirely.
When Google released the consumer Gemini 2.5 Deep Think a week and a half later, it was explicit that this was a different, faster variant. The publicly available product reasons in seconds or minutes rather than the hours the research model needed, and internally it reaches only bronze-level performance on the 2025 IMO benchmark. [2][3] The gold-medal research version was instead shared with a small group of mathematicians and academics for testing.
The announcement also drew attention because of its timing. OpenAI publicized a comparable result, also 35 points and five problems solved, on July 19, two days before Google DeepMind. [6] Demis Hassabis said the company had waited so that it could honor the IMO board's request that AI labs hold their results until after official grading and after human contestants received recognition. [6]
Google reported state-of-the-art results for the consumer Deep Think on competition coding and on Humanity's Last Exam, a broad test of expert-level knowledge across science and mathematics. [2] The figures below were cited at the August 2025 launch.
| Benchmark | Gemini 2.5 Deep Think | Comparison |
|---|---|---|
| Humanity's Last Exam | 34.8% | Grok 4: 25.4%; OpenAI o3: 20.3% [3] |
| LiveCodeBench V6 | 87.6% | Grok 4: 79%; OpenAI o3: 72% [3] |
| MMMU (multimodal, I/O preview) | 84.0% | reported at I/O, May 2025 [1] |
The LiveCodeBench score improved from roughly 80.4% at the May preview to 87.6% at the August release, which Google credited to further research and tester feedback. [3] At I/O, Google also reported strong results on the 2025 USAMO, which it called one of the hardest math benchmarks then in use. [1] The model accepts up to about 1 million input tokens, consistent with the Gemini 2.5 family's long-context design.
Gemini 2.5 Deep Think is available to subscribers of Google AI Ultra, the company's top consumer plan priced at $249.99 per month. [2][3] There is no per-query surcharge, but usage is capped: subscribers get only a fixed, small number of Deep Think prompts per day. [2][3] Users access it by selecting Gemini 2.5 Pro in the model menu and toggling Deep Think in the prompt bar.
Google said developer access would follow within weeks, offering the model through the Gemini API in two forms, with and without tools, initially for trusted testers and enterprise evaluation. [2][3] The gold-medal research model was not made part of either the consumer or general API offering.
Coverage framed Deep Think as a notable advance in test-time reasoning, with TechCrunch, VentureBeat, and others emphasizing its multi-agent, parallel approach and its IMO pedigree. [3] At the same time, reporting stressed the gap between the headline gold-medal achievement and what subscribers actually receive, since the consumer version is a faster, less capable variant and is rationed to a handful of prompts a day on a $250 plan. [2][3] Google noted that Deep Think showed improved content safety relative to Gemini 2.5 Pro, while also exhibiting a higher tendency to refuse some benign requests. [2] Deep Think is distinct from Google's separate research efforts in formal mathematics such as AlphaProof and AlphaGeometry, and it is offered alongside Gemini's other tiers, including the consumer Gemini Advanced experience.