Gemini 2.5: Gemini makes thinking explicit

The announcement and the narrow claim

On 25 March 2025, Google DeepMind announced Gemini 2.5 Pro Experimental and described it in one load-bearing sentence: Gemini 2.5 models are thinking models, meaning they are built to reason through a problem internally before producing an answer. The company said the experimental model debuted at the top of the LMArena human-preference leaderboard, and it shipped first through Google AI Studio and the Gemini app for Advanced subscribers, with Vertex AI access promised in the following weeks. That is the documented event: a specific model, on a specific date, presented under a specific framing.

It helps to keep two levels of claim apart. The narrow claim is precise and verifiable: Google released a reasoning-oriented, multimodal, long-context model and reported a set of benchmark results for it. The broad historical judgment, that inference-time reasoning became the dominant product direction across major providers in 2025, is a separate assessment that depends on what competitors were doing and on what followed. Conflating the two turns a well-documented launch into an inflated origin story. This report treats Gemini 2.5 as an important consolidation of an already-emerging idea rather than as the invention of reasoning in language models.

What "thinking model" actually names

The phrase "thinking model" is a product label for a technical shift in where a system spends its effort. A conventional autoregressive model answers a prompt directly, so its quality is fixed once training ends. A thinking model is trained and configured to generate an extended internal reasoning process first, then a final response, spending additional computation at inference time on hard problems. Google framed this as reasoning capabilities now built directly into its models rather than bolted on as a separate feature.

The intellectual lineage is older than the label. Chain-of-thought prompting, described by Wei and colleagues in 2022, showed that simply asking a model to produce intermediate reasoning steps improved performance on arithmetic, commonsense and symbolic tasks. That was a prompting technique applied to fixed models. The 2024 to 2025 generation of thinking models internalized the behaviour: instead of relying on the user to request steps, the model is trained to produce them by default and to use more of them when a problem is harder. Gemini 2.5 belongs to that second phase.

The reasoning turn that preceded it

Gemini 2.5 did not arrive into an empty field. The immediate precedent was OpenAI, which in September 2024 introduced its o1 series and stated that the model learns to reason using a long internal chain of thought trained with reinforcement learning, and that its performance improved with more reinforcement-learning training and with more time spent thinking at test time. That was the clearest public statement that a distinct product category, the reasoning model, had formed.

The second precedent was open. In January 2025, DeepSeek released a paper reporting that reasoning ability could be incentivized through pure reinforcement learning without human-labelled reasoning traces, and it published open weights for DeepSeek-R1. That combination, a strong reasoning model with visible methodology and downloadable weights, pushed every major lab to demonstrate comparable capability. Seen against these two events, Gemini 2.5 Pro Experimental is best read as Google's answer within a competition already underway: the constraint it addressed was not a missing idea but a missing top-tier Google entry in the reasoning category.

From Gemini 1.0 to 2.5: the family lineage

The 2.5 release also sits inside a fast-moving model family. Google introduced the line in December 2023 with Gemini, described as a family of natively multimodal models trained jointly across text, images, audio and video rather than stitched together from separate encoders. That native multimodality is why Gemini 2.5 could accept text, audio, images, video and whole code repositories as input without a bolted-on adapter.

The second inherited capability was long context. In early 2024, Google described Gemini 1.5 as unlocking understanding across millions of tokens of context, using a mixture-of-experts design and demonstrating retrieval across very long inputs. Gemini 2.5 Pro Experimental launched with a one-million-token context window and a two-million-token window promised, so the long-context work was a direct predecessor rather than a new contribution of the 2.5 launch. Reasoning was the newer ingredient: Google had shipped an earlier Gemini 2.0 Flash Thinking experiment in December 2024, and 2.5 folded that thinking behaviour into a flagship Pro model alongside the existing multimodality and long context.

How inference-time reasoning works

Mechanically, a thinking model changes the compute budget rather than the underlying architecture. The Transformer backbone remains; what differs is that the model is trained to emit a long intermediate reasoning trajectory and is permitted to consume more tokens, and therefore more computation, on difficult inputs. The research foundation for that trade is explicit. Snell and colleagues argued in 2024 that scaling test-time compute can be more effective than scaling model parameters, finding that on problems where a smaller base model has a non-trivial success rate, allocating more inference-time computation could outperform a much larger model under a matched compute budget.

Two training ingredients make that trajectory useful rather than merely long. The first is reinforcement learning against outcomes, so the model is rewarded for reasoning paths that reach correct answers, the approach OpenAI and DeepSeek both described. The second is verification of intermediate steps. OpenAI researchers had earlier shown in Let's Verify Step by Step that supervising each step of a solution, not just the final answer, produced more reliable reasoning on mathematics problems. Google did not publish the exact recipe for the experimental 2.5 model at launch, so the mechanism is stated here at the level the field documents, not as a claim about Google's specific internal method.

The benchmark evidence, read carefully

Google reported concrete numbers, and they should be read as task-specific results, not as a general intelligence score. On Humanity's Last Exam, a deliberately hard multi-domain exam built by hundreds of subject experts, Google reported that Gemini 2.5 Pro reached 18.8 percent without tool use, which it presented as a state-of-the-art result at the time. That figure is informative precisely because the benchmark was designed so that leading models score low: a number under 20 percent is the point, illustrating how far frontier reasoning still had to go.

On coding, Google reported 63.8 percent on SWE-bench Verified, a benchmark of real GitHub issues, and it stated the result used a custom agent setup, which matters because agent scaffolding materially affects the score. It also claimed state-of-the-art results on GPQA, a graduate-level science question set, and on the AIME 2025 mathematics competition without majority-voting tricks. The headline positioning came from LMArena, the crowd-sourced human-preference platform, where Google said the model led by a significant margin. A later preview update in May 2025 reported the model at number one on the WebDev Arena leaderboard for building web apps and a score of 84.8 percent on the VideoMME video-understanding benchmark, so some of the strongest coding claims belong to iterations after the March launch, not to the experimental release itself.

Competing approaches of the moment

The reasoning category in early 2025 contained genuinely different strategies, and the comparison is more instructive than a ranking. OpenAI's o-series kept the internal chain of thought hidden from the user and sold access to a reasoning tier as a distinct product. DeepSeek's R1 exposed its reasoning traces and released open weights, betting on transparency and distribution. Google's move with 2.5 was to integrate reasoning into a single flagship that also carried native multimodality and a very large context window, positioning it as a general workhorse rather than a specialized reasoning endpoint.

These are contrasting product philosophies, not just different scores. A hidden reasoning trace protects a lab's methodology and reduces certain misuse risks, while an exposed one aids research and trust. Pure reinforcement learning without human reasoning labels, as DeepSeek reported, lowers the cost of building a reasoning model but raises questions about reliability and readability of the resulting traces. Placing Gemini 2.5 next to these live alternatives, rather than against an outdated non-reasoning baseline, is the honest way to judge what the March 2025 release actually offered.

Distribution as strategy, not a footnote

Where a model ships is part of what the event was. Gemini 2.5 Pro Experimental appeared first in Google AI Studio, the developer surface, and in the consumer Gemini app for Advanced subscribers, with enterprise access through Vertex AI following, and Google said pricing would be introduced in the following weeks. That staged rollout, developer and consumer first, priced enterprise access later, reflects a company able to distribute a frontier model across a search, cloud and device footprint that most competitors lacked.

The word "experimental" in the name is not decoration. It signalled an early, rapidly iterating release rather than a stable production model with fixed behaviour and guarantees. The grounding record is careful about this: the experimental launch differs from the later stable variants, and provider-reported benchmarks are evidence from the vendor, not independent proof of reliable reasoning on every task. The staged, experimental distribution is why the March 2025 event is best cited as a launch of a direction, with the stabilized Gemini 2.5 Pro and Flash models that followed being the productized outcome.

The documented downstream influence

The clearest downstream evidence is Google's own consolidation of the family. In July 2025, the Gemini team published a technical report titled Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities, which documents the family as a coherent set of reasoning-first models rather than a single experiment. That report is the point where the March experimental claim became a stated, versioned capability with a full author list behind it.

The broader influence is the normalization of user-visible thinking. Through 2025, reasoning modes, thinking toggles and reasoning-effort settings became standard across major assistants and APIs, and the idea that a model should spend variable computation per query moved from a research finding into an ordinary product control. Gemini 2.5 did not originate that pattern, OpenAI's o1 and DeepSeek's R1 preceded it, but a flagship from Google adopting the same framing helped settle the industry consensus that a serious general model in 2025 was expected to reason before answering. That is a defensible influence claim because it rests on observable product convergence, not on speculation about internal methods.

Honest limits and caveats

Several caveats bound the significance of this event. First, the numbers are provider-reported and configuration-dependent: the SWE-bench figure relied on a custom agent, and benchmark scores in this period were sensitive to scaffolding, prompting and evaluation choices, so cross-vendor comparisons require matching those conditions. Second, a strong reasoning benchmark does not establish reliable reasoning in general; models that produce fluent step-by-step traces can still reach wrong conclusions confidently, and a chain of thought is not by itself a guarantee that the model actually followed that logic to its answer.

Third, inference-time reasoning has real costs. Generating long internal traces increases latency and token consumption, which is why reasoning is typically reserved for harder queries rather than applied uniformly, and why the economics of a thinking model differ from a direct-answer model. Fourth, the experimental label means the March 2025 behaviour was provisional; specific results and quirks of the experimental checkpoint should not be treated as permanent properties of the Gemini 2.5 line. None of these caveats diminishes that a leading lab shipped a competitive reasoning model; they simply keep the claim proportional to the evidence.

How to place Gemini 2.5 in AI history

The most accurate placement is as a consolidation point, not a discovery. The narrow, well-sourced fact is that on 25 March 2025 Google DeepMind released a multimodal, long-context, reasoning-oriented model and reported leading results on human-preference and reasoning benchmarks. The wider judgment, that 2025 was the year inference-time reasoning became a default expectation for frontier models, is supported by the surrounding record of o1, DeepSeek-R1 and the rapid spread of thinking modes, and Gemini 2.5 is one of the events that makes that trend legible rather than the single cause of it.

A reader navigating the timeline can use three questions. Which earlier idea supplied the necessary component? Here, chain-of-thought prompting and test-time-compute scaling supplied the reasoning idea, and Gemini 1.0 and 1.5 supplied native multimodality and long context. Which contemporaries pursued the same goal differently? OpenAI with hidden traces and DeepSeek with open weights. Which limitation still required further work? Reliability, cost and independent verification of reasoning quality. Answering those keeps Gemini 2.5 in proportion: a strong, well-distributed entry that helped standardize explicit reasoning, inheriting more than it invented and leaving the hardest reliability questions open for the work that followed.

References

• Google DeepMind (2025). Gemini 2.5: Our most intelligent AI model.
• Google Developers (2025). Gemini 2.5 Pro Preview: even better coding performance.
• Gemini Team, Google DeepMind (2025). Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities.
• OpenAI (2024). Learning to Reason with LLMs.
• DeepSeek-AI (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.
• Wei et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.
• Snell et al. (2024). Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters.
• Lightman et al. (2023). Let's Verify Step by Step.
• Gemini Team, Google (2023). Gemini: A Family of Highly Capable Multimodal Models.
• Gemini Team, Google (2024). Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.
• Rein et al. (2023). GPQA: A Graduate-Level Google-Proof Q&A Benchmark.
• Jimenez et al. (2023). SWE-bench: Can Language Models Resolve Real-World GitHub Issues?.
• Phan et al. (2025). Humanity's Last Exam.
• Chiang et al. (2024). Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference.