Gemini

Gemini is the family of multimodal models developed by Google DeepMind, announced on 6 December 2023 as the successor to the company's earlier PaLM and LaMDA lines. It was trained on text, images, audio and video together rather than assembled from separate systems, and it was released in three sizes at once: Ultra, Pro and Nano 1.

Gemini is unusual among model families in how deeply it is embedded in its owner's products. The same technology powers the Gemini app, AI Overviews in Search, assistance in Gmail and Docs, features on Android, and the developer API and Vertex AI platform. That distribution gives Google a reach no stand-alone lab can match, and it also means that changes to the models surface simultaneously across products used by billions of people 2.

This entry covers the family's structure and release history, what is known about its training, its capabilities and limits, its benchmark record, pricing and access, and the disputes over demonstrations, image generation and search accuracy that have shaped its reception.

Gemini models are sold in tiers that have changed names repeatedly. The original release used Ultra, Pro and Nano; later generations used Pro, Flash, Flash-Lite and Nano, with Pro models positioned for reasoning and Flash models for speed and cost. The naming has not been stable, which makes version histories difficult to follow and gives the misleading impression of a single product line with a fixed hierarchy 2.

The models are natively multimodal. They accept text, images, audio and video as input and can produce text, and some versions produce images and speech through connected models. Long context is a signature feature: the 1.5 generation introduced a one-million-token window, and later Pro models retained it, enough for a feature film, a large codebase or several hundred pages of documents in one request 3.

Gemini also comes in on-device sizes. Nano models run locally on phones and other hardware for tasks such as summarisation and suggestion, where sending data to a data centre would be slow or inappropriate. Google has presented this range — from on-device to data-centre scale — as the point of the family, though the small models are considerably less capable than the large ones 1.

The product layer is broader than the model. The Gemini app handles conversation, image generation, research and voice; Workspace puts Gemini into mail, documents and spreadsheets; Android exposes it as a system assistant; and Vertex AI sells it to enterprises. Each surface has its own terms, limits and data handling 2.

Gemini 1.0 was announced in December 2023, with Ultra available to enterprise customers and Pro and Nano released more widely. Google reported that Ultra was the first model to exceed human-expert performance on MMLU, a multi-subject knowledge test, at 90.0 per cent. The launch was accompanied by a demonstration video that Google later acknowledged had been edited for pace 1.

Gemini 1.5 Pro followed in February 2024 with a mixture-of-experts architecture and a context window of up to one million tokens, extended to two million for testing. Gemini 1.5 Flash, a smaller and cheaper variant, arrived in May 2024. The 1.5 generation was also when Google began reporting near-perfect recall on long-document retrieval tests, a capability that became central to its marketing 3.

Gemini 2.0, released in December 2024, shifted the emphasis from conversation to agents: native tool use, multimodal output and integration with Google's services. Gemini 2.5 Pro in March 2025 introduced a thinking mode and topped several public leaderboards on release, and the 2.5 generation was extended through Flash and Flash-Lite versions during the year. A third generation, Gemini 3, was released in late 2025 with further gains claimed in reasoning and coding 4 5.

Each generation has been followed by point releases and by retirements of older models. Google publishes deprecation schedules for the API, and some consumer features move between models without notice. For developers the practical consequence is that a pinned model version is a temporary commitment rather than a permanent one 6.

Google has published more about Gemini's training than some competitors, though not enough to reproduce it. The technical reports describe transformer-based models with mixture-of-experts layers in the 1.5 generation, trained on a multimodal mixture of text, images, audio and video, and served on tensor processing units designed in-house. Parameter counts are not disclosed for the frontier models 3.

Training runs on TPU pods rather than on the GPUs that most other labs use. Google designs both the chips and the models, which allows co-design of the hardware and the training recipe and gives the company a supply advantage that does not depend on external accelerator vendors. The reports describe training across thousands of chips and the infrastructure needed to keep them fed 4.

The 2.5 generation added reinforcement learning for reasoning. The models are trained to produce a longer internal deliberation before answering, and the amount of computation can be varied at inference. Google has also described distilling larger models into smaller ones, so that the Flash models inherit some capability from the Pro models rather than being trained from scratch 4.

Safety work is described in model cards rather than in the technical reports. Google evaluates Gemini for harmful content, bias, cyber and biological uplift, and reports the results to internal review bodies; the company's AI principles list applications it says it will not pursue, including weapons and mass surveillance. How strictly those principles bind particular deployments is a question the company answers case by case 7.

Gemini's most distinctive capability is breadth of input. It can analyse a video, read a scanned document, transcribe audio and answer questions that span all three, which makes it useful for media archives, compliance review and accessibility work. Its long context window is genuinely usable for retrieval over large document sets, though attention over a million tokens is uneven and important details late in a prompt can be missed 3.

Coding and reasoning are competitive with the frontier. The Pro models handle multi-file code changes, debugging and explanation; the thinking variants improve results on mathematics and science at the cost of latency. Integration with Google's own tooling — Search, Maps, YouTube, Workspace — gives the assistant access to live information that a stand-alone chatbot lacks 4.

The limits are familiar. Gemini fabricates facts and citations, and its errors are more consequential when they appear inside Search summaries or a user's email. Image generation, which runs through a separate model, has been prone to historical and demographic inaccuracies that Google has struggled to calibrate. On-device Nano models are useful for narrow tasks but cannot follow complex instructions 5.

Availability is itself a limit. Some features launch in the United States first, and European releases have been delayed by regulatory review. Enterprise customers face additional constraints on data residency and on which models can be used with which Workspace data. The result is a family whose capability depends heavily on where and how it is accessed 2.

Google reports benchmark results at every launch, and the figures are vendor-reported. Gemini 1.0 Ultra was reported at 90.0 per cent on MMLU, 62.4 per cent on MMMU and 94.4 per cent on GSM8K. Gemini 1.5 Pro was benchmarked extensively on long-context retrieval, multimodal understanding and code, and the 2.5 generation added results on GPQA, AIME, SWE-bench Verified and agentic tool-use suites 1 4.

Comparison is difficult because Google has often benchmarked a new model against its own previous generation rather than against the strongest competitor. That is a reasonable way to show progress and a poor way to establish leadership, and critics have noted that the choice of baseline can change the story. The company has also reported results with and without tools, which further complicates like-for-like comparison 5.

The benchmarks themselves are weakening. MMLU is saturated, GPQA is widely circulated, and coding benchmarks depend heavily on the harness. Google has responded by commissioning external evaluations and by reporting results on newer, harder tests, but those tests are usually designed by the same community that builds the models, and their useful life is short 4.

Independent evaluation has been mixed. Third-party leaderboards have generally placed Gemini models near the top, particularly on long-context and multimodal tasks, while replicating fewer of the headline margins. The pattern is consistent with the rest of the industry: the models are genuinely strong, and the specific numbers in the launch post are the least reliable part of the announcement 5.

The Gemini app is free with usage limits. Google AI Pro, a subscription bundled with additional storage and Workspace features, raises those limits, and Google AI Ultra, priced far higher, adds access to the largest models and to advanced video and agent features. The plans have been renamed and repriced repeatedly, and some features come with hardware purchases or enterprise agreements rather than being sold separately 6.

On the API, Google prices per million tokens with separate rates for input, output and cached input, and a free tier for low-volume use. At the 2.5 generation, Pro was priced at 1.25 US dollars per million input tokens and 10 dollars per million output tokens for prompts under 200,000 tokens, with higher rates above that threshold; Flash and Flash-Lite were substantially cheaper. Long-context requests therefore cost more in two ways, through more tokens and through a higher per-token rate 6.

Enterprises buy through Vertex AI, which adds controls for data residency, access management and auditing, and through Workspace agreements that bundle Gemini into productivity software. Google also offers Gemini through its cloud marketplace and as part of Android and Chrome integrations. The commercial model is less a single price than a set of bundles, which makes direct comparison with rivals difficult 2.

Free access is part of the strategy. Google can afford to give away inference because the models run on its own TPUs and because the assistant feeds its advertising and cloud businesses. That cross-subsidy is a structural advantage, and it is also why pricing in this market has fallen faster than a pure cost analysis would predict 6.

The first controversy concerned the launch demonstration. The December 2023 video showed Gemini responding to speech and images in what appeared to be real time; Google later confirmed that the sequences had been sped up and that some responses were produced from still images rather than live video. The company described the edits as a way to make the demo digestible, and critics described them as misleading 5.

The second concerned image generation. In February 2024 users found that Gemini's image tool was producing historically inaccurate depictions — including generated images of diverse Nazi soldiers — after attempts to broaden representation. Google paused generation of people, apologised, and rebuilt the feature. The episode became a case study in the difficulty of specifying demographic behaviour in a generative system, and it was criticised from both directions, for bias and for overcorrection 5.

The third concerned Search. AI Overviews, which place Gemini-generated summaries above conventional results, produced a series of documented errors after its May 2024 rollout, including advice to eat rocks and to add glue to pizza. Google called the examples uncommon and adjusted the system, but the episode illustrated the risk of putting a probabilistic model in a position of factual authority 5.

Regulatory delay is a fourth theme. The Gemini app reached the European Union later than the United States, and features such as image generation and connected apps have arrived in different markets at different times. Google attributes the sequencing to local law and review; critics see a company large enough to shape the rules choosing where to comply first 7.

The central dispute is whether Gemini leads the field. Google has claimed state-of-the-art results at several launches, and some independent leaderboards have supported the claim on release, particularly for long-context and multimodal tasks. Other evaluators have found that the lead is narrow, that it does not survive changes in prompting or harness, and that competitors release comparable models within weeks. The honest summary is that Gemini is at or near the frontier, not that it is reliably first 4.

Safety claims are contested in a different way. Google publishes model cards and says it applies its AI principles before deployment, and it has pointed to red-teaming and to external audits. Critics note that the company removed its earlier pledge not to use AI for weapons and surveillance, that its principles are self-imposed, and that a model embedded in Search reaches people who never chose an AI system. Google says the principles are living documents and decisions are reviewed case by case 7.

The demonstration controversy still colours coverage. Google's explanation — that the video was edited for time and that the underlying capabilities were real — is plausible and was partly borne out by later products. But the episode established a pattern in which launch material is treated as advertising rather than evidence, and it made journalists and researchers more sceptical of subsequent Gemini announcements 5.

What is not disputed is the distribution. Gemini is embedded by default in a search engine, a mobile operating system, an office suite and a cloud platform used by billions of people. That reach is the strongest argument for its importance and the strongest reason to scrutinise the claims made for it: a model deployed at that scale does not need to be the best to shape what most people encounter as artificial intelligence 2.