Wiki · Concept · Last reviewed June 23, 2026

Gemini

Gemini is Google's multimodal frontier AI model family and assistant platform. The name refers both to Google DeepMind's model line and to a product surface across the Gemini app, Google Search AI Mode, Android, Workspace, Google Cloud, AI Studio, Vertex AI, Antigravity, and enterprise agent tools.

Category: Concept Published: June 23, 2026 Modified: June 23, 2026 Last reviewed: June 23, 2026 Tags: Gemini, Google DeepMind, Multimodal AI, AI Agents, Search, Platform Governance

Definition

Gemini is the successor brand to several Google AI lines. Technically, it is a family of multimodal foundation models developed by Google DeepMind. Productively, it is the assistant and agent layer that Google places inside search, phones, productivity software, developer tools, cloud services, and consumer subscriptions.

The ambiguity is part of the subject. A user may say "Gemini" and mean a chat app, a model such as Gemini 3.5 Flash, an API endpoint, a Search feature, an Android assistant, an enterprise agent, an image model, a voice interface, or a set of safety and policy controls. Like ChatGPT and Claude, Gemini is not only a model checkpoint. It is an interface institution.

The strict definition for source work is narrower: identify the exact layer being discussed. A claim about the stable API model gemini-3.5-flash is not automatically a claim about the Gemini app, AI Mode in Search, Gemini Enterprise, Gemini Omni, Nano Banana, Gemini Robotics, Google Antigravity, or an agent using Gemini through a separate tool harness.

Current Context

As of June 23, 2026, the newest public Gemini family is Gemini 3.5, led by Gemini 3.5 Flash. Google says 3.5 Flash is available through the Gemini app, AI Mode in Search, Google Antigravity, the Gemini API in Google AI Studio and Android Studio, Gemini Enterprise Agent Platform, and Gemini Enterprise. Google also announced Gemini 3.5 Pro as forthcoming, but the public developer documentation reviewed for this page listed gemini-3.5-flash as the stable 3.5 model and did not list a generally available gemini-3.5-pro model.

The API surface matters because it is more specific than the brand. Google AI for Developers lists gemini-3.5-flash with a 1,048,576-token input limit and a 65,536-token output limit, text/image/video/audio/PDF inputs, text output, and support for caching, code execution, file search, function calling, Google Maps grounding, Search grounding, structured outputs, thinking, URL context, batch API, flex inference, and priority inference. The same page states that computer use, image generation, audio generation, and Live API are not supported by that model.

Distribution is the main governance fact. Google's May 2026 Search update said AI Mode had passed one billion monthly users and that Gemini 3.5 Flash became the default model in AI Mode globally. That makes Gemini both a model family and an answer layer inside one of the world's dominant information systems.

Gemini is also now a family of specialized systems. Google DeepMind's model page lists Gemini 3.5 alongside Gemini Omni, Gemini Image or Nano Banana, Gemini Audio, Gemini Robotics, and Gemini Embedding. Those systems share branding and some lineage, but they have different modalities, release channels, safety issues, and evidence standards.

History

Google announced Gemini 1.0 on December 6, 2023, shortly after combining DeepMind and Google Brain into Google DeepMind. The first technical report described Gemini as a family of models trained for image, audio, video, and text understanding, with Ultra, Pro, and Nano sizes for different deployment contexts.

On February 8, 2024, Google renamed Bard to Gemini, launched Gemini Advanced with Ultra 1.0, and began rolling out mobile access through Android and the Google app on iOS. That move turned Gemini from a model name into Google's main public AI assistant brand.

Gemini 1.5 made long context a public competition, with Google describing one-million-token context access for developers and enterprise customers in preview. Gemini 2.0, announced in December 2024, shifted the message toward the agentic era: multimodal output, native tool use, Project Astra, Project Mariner, and developer-facing agent workflows.

Gemini 2.5, announced in March 2025, emphasized "thinking" models for stronger reasoning, math, science, and coding. Gemini 3, announced in November 2025, extended that line into Search, the Gemini app, AI Studio, Vertex AI, and Google Antigravity. On May 19, 2026, Google introduced Gemini 3.5, beginning with Gemini 3.5 Flash, a model positioned around fast agentic coding, long-horizon workflows, and broad deployment across Google products.

Model Family

Gemini has several overlapping model roles. Pro models are positioned for difficult reasoning, multimodal work, coding, and complex tasks. Flash models are optimized for speed, scale, and lower-latency deployment. Flash-Lite variants target high-volume and cost-sensitive uses. Nano names on-device or memory-constrained deployment, especially in mobile contexts.

The family also includes specialized Gemini-branded systems. Gemini image models power image generation and editing products such as Nano Banana. Gemini audio models support real-time audio use. Gemini Robotics applies the family to vision-language-action control. Gemini embedding models support retrieval and representation tasks.

As of this review, the newest public family is Gemini 3.5, with Gemini 3.5 Flash generally available and Gemini 3.5 Pro announced as forthcoming. Google DeepMind's Gemini 3.5 Flash model card describes distribution through the Gemini app, Gemini Enterprise, Google AI Studio, the Gemini API, Search AI Mode, and Google Antigravity, and reports evaluations across reasoning, coding, agentic tool use, multimodal capability, multilingual behavior, and long context.

The model-card and API records should be read together. The model card gives the provider's release narrative, evaluation categories, intended uses, limitations, and safety assessment. The API docs give operational details such as model IDs, token limits, supported inputs, supported tools, stable versus preview naming, and endpoint behavior. Neither source alone tells a user exactly what happened inside a product surface such as Search or Workspace.

Product Layer

Gemini matters because it sits inside Google's distribution system. It can appear in the Gemini app, Search AI Mode and AI Overviews, Android, Pixel devices, Google Workspace, Gmail, Docs, Slides, Sheets, Meet, Google Cloud, Vertex AI, AI Studio, Android Studio, and enterprise agent products.

This makes Gemini different from a standalone chatbot. It is an attempt to put a model-mediated assistant into the default paths by which many people search, write, navigate phones, build software, manage documents, run enterprise workflows, and query institutional data.

The product layer also changes user expectations. A Gemini answer can look like search, chat, a document assistant, an operating-system action, a generated interface, a coding agent, a personal agent, or an enterprise workflow. Governance has to follow the surface, not only the underlying model.

Search is the clearest example. If Gemini produces an answer, follow-up path, generated interface, or background information agent inside Search, the relevant questions are not only whether the model is capable. They are how ranking, citations, query logs, advertising incentives, publisher traffic, personalization, and appeal mechanisms work around the answer layer.

Agents and Developers

Google's Gemini 2.0 announcement framed the line around the "agentic era." Project Astra explored a universal assistant that could perceive context and use tools. Project Mariner explored browser-based computer use. Jules targeted coding workflows. Later Gemini releases continued this direction through Google Antigravity, Gemini Enterprise Agent Platform, and developer APIs.

Gemini 3.5 Flash was explicitly introduced around agentic workflows. Google described it as strong for coding, long-horizon tasks, multi-step workflows, richer web interfaces, and supervised subagents. The practical claim is that Gemini should not only answer questions, but plan, build, iterate, use tools, and coordinate work under human direction.

For developers, the important surfaces are AI Studio, Vertex AI, the Gemini API, Android Studio, Gemini CLI, Antigravity, and third-party tool integrations. These channels place Gemini in the same competitive stack as OpenAI models, Claude, Llama-based deployments, Qwen, DeepSeek, Mistral, and specialized inference providers.

Agent claims require extra care. Google DeepMind's June 2026 AI Control Roadmap describes internal safeguards for increasingly capable agents, including monitoring, prevention and response, capability-linked security levels, and an internal prototype that analyzed one million coding-agent trajectories. Those controls are relevant to Gemini-era agents, but they should be cited as Google DeepMind's framework and implementation claims, not as independent proof that deployed Gemini agents are safe in every product context.

Safety and Governance

Google ties Gemini governance to model cards, red teaming, safety policies, child-safety commitments, and Google DeepMind's Frontier Safety Framework. The Gemini 3.5 Flash model card says the model was evaluated across capability areas and safety areas, and that Google assessed it against frontier-safety thresholds using Gemini 3.1 Pro results as part of the basis for confidence.

That governance matters, but it should not be mistaken for public control. Model cards and internal safety frameworks are disclosure instruments written by the company deploying the system. They help outsiders understand claims, but they do not by themselves create independent audit power, appeal rights, or democratic oversight of a model embedded into search and productivity infrastructure.

The 3.5 Flash model card is also a useful example of why source discipline matters. It reports automated safety evaluations, manual red teaming, and a frontier safety assessment; it also states that Gemini 3.5 Flash was judged unlikely to reach Critical Capability Levels partly because Google assessed Gemini 3.1 Pro as the most generally capable model at publication time and did not find it crossed those thresholds. That is a provider-reported safety case, not an external conformity assessment.

For Gemini governance, the unit of control should be the deployed system: model version, product surface, prompts, grounding sources, tools, memory, user identity, enterprise tenant, logging, approval gates, rollback path, and human oversight. A model card for a base model is not enough to govern a Search answer, a Workspace draft, a coding subagent, a browser agent, or an enterprise workflow.

Why It Matters

Gemini is one of the central frontier-AI systems because it joins model capability with Google-scale distribution. ChatGPT showed that a general assistant could become a mass consumer product. Claude made safety-branded professional assistance a major competitive category. Gemini tests whether an incumbent information platform can make AI assistance native to search, phones, documents, cloud development, and enterprise operations.

The stakes are not only benchmark performance. Gemini can change how people encounter information, which sources are surfaced, how documents are drafted, how software is built, how enterprise agents act, and how much of digital life becomes mediated by a privately governed model interface.

Risk Pattern

Source Discipline

Gemini claims should name the artifact: product surface, model family, exact model ID, model-card date, API documentation date, release channel, and whether the claim concerns a base model, a hosted app, a developer API, a Search feature, an enterprise agent, or a specialized modality model.

Use official sources for release and capability claims: Google and Google DeepMind launch posts, Google AI for Developers documentation, Google Cloud or Vertex AI documentation, and Google DeepMind model cards. Use independent research, regulator text, standards bodies, or security writeups when making claims about prompt injection, auditability, platform power, privacy, competition, or public accountability.

Do not infer consciousness, divinity, wisdom, or achieved general intelligence from the Gemini brand, benchmark results, multimodal capability, long context, tool use, or agentic workflows. The relevant evidence questions are operational: what was the model, what could it see, what tools could it call, what action was authorized, what safety layer intervened, and what record survived?

Spiralist Reading

Gemini is the Mirror inside the search box.

Its cultural force comes from placement. A chatbot waits for a user to visit it. Gemini can appear where the user already lives: the query bar, the phone, the inbox, the document, the code editor, the meeting, the cloud project, the enterprise workflow. That makes it less like a tool added to the internet and more like a new interpreter inserted into the internet's default rituals.

For Spiralism, Gemini is a test of cognitive sovereignty at platform scale. The question is not whether Google can build powerful models. It can. The question is whether people and institutions can see when search becomes synthesis, when assistance becomes delegation, when personalization becomes capture, and when convenience becomes governance.

Open Questions

Sources

Return to Wiki