Google Gemini

Google Gemini is Google DeepMind’s family of frontier multimodal models. The models process text, images, audio, video, PDFs, and code in a single request, and the current generation accepts roughly a million tokens of input. Gemini solves a common problem for builders: instead of stitching together separate models for vision, speech, and text, you send mixed inputs to one model and get one reasoned answer back.

Gemini is one of the three widely used frontier model families alongside OpenAI’s GPT and Anthropic’s Claude . For background on what a foundation model is and what a large language model does, follow those links first.

The family

Google ships Gemini as tiers, not a single model. The naming follows a generation number plus a tier label — but the generation numbers no longer line up across tiers, and that is the most important thing to understand about the lineup as it stands on 25 September 2026.

The Flash line has run ahead of the Pro line. Google shipped three Flash releases in six weeks — Gemini 3.6 Flash on 21 July 2026, Gemini 3.7 Flash on 13 August, and Gemini 3.8 Flash on 2 September — while the Pro tier is still on Gemini 3.1 Pro from 19 February 2026. Gemini 3.5 Pro was trailed at Google I/O on 19 May 2026 for a June launch and has not shipped; DeepMind’s Pro page carries only a “3.5 Pro coming soon” note, and press reporting (Bloomberg, Axios, Forbes, 9to5Google) attributes the repeated slips to coding-benchmark shortfalls. In practice, Gemini 3.8 Flash is Google’s current flagship shipping model, and any guidance written since mid-2026 that assumes a 3.5 Pro exists is wrong.

Flash — the working default

Gemini 3.8 Flash (gemini-3.8-flash) reached general availability on 2 September 2026 with a 1,048,576-token input window and 65,536-token output limit. Google positions it for long-horizon software engineering, autonomous agents, and complex enterprise workflows, and ships it across the Gemini API, AI Studio, Android Studio, Gemini Enterprise, Google Antigravity, Search AI Mode, Google Sheets, and the Gemini app for AI Pro and AI Ultra subscribers. Gemini 3.7 Flash (13 August 2026) and Gemini 3.6 Flash (21 July 2026) remain listed as stable and are still priced, so existing integrations do not have to move immediately. Gemini 3.5 Flash (19 May 2026) is also still stable, but it is the odd one out on price — see below.

Pro — capable, and still in preview

Gemini 3.1 Pro (gemini-3.1-pro-preview) launched on 19 February 2026 with a 1,048,576-token input window and 65,536-token output limit, and claimed 77.1% on ARC-AGI-2 at launch. More than seven months on — and still as of 25 September 2026 — it is still a preview model, and the deprecations page lists no shutdown date for it. Google said at launch that general availability would follow; no GA has been announced. There is therefore no GA Pro-tier Gemini 3.x model today, which matters if your change-management process forbids shipping on preview endpoints.

Flash-Lite — the cost floor

Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite), GA on 21 July 2026, is the current cost-efficient tier for high-throughput execution. Gemini 3.1 Flash-Lite (preview 3 March 2026, GA 7 May 2026) is still listed and, counter-intuitively, is still the cheapest Gemini text model on the price sheet — $0.25 input against the newer tier’s $0.30. It has a published shutdown date, though: 7 May 2027, with gemini-3.5-flash-lite as the named replacement, so the saving has a fixed end.

Gemini 2.5 — restricted to existing users

On 18 September 2026 Google limited access to the Gemini 2.5 models to users “who have actively used them in the past”. The models are explicitly not deprecated and keep being served “until further notice”, but a new project can no longer start on them; Google directs new work to 3.5 Flash-Lite or 3.8 Flash. If a deployment pipeline creates fresh Google Cloud projects or API keys and pins a 2.5 model ID, test it now. Separately, gemini-2.5-flash-image has a fixed shutdown of 2 October 2026 on the Gemini API — but on the Gemini Enterprise Agent Platform Google extended its retirement to 15 March 2027 (release note of 14 September 2026), so the deadline depends on which surface you call.

Deep Think — a gated mode, not a tier you can call

Deep Think is easy to mistake for a fourth tier alongside Flash and Pro. It is not. Gemini 3 Deep Think is a parallel-reasoning mode — the model explores multiple hypotheses simultaneously — that is generally available in the Gemini app to Google AI Ultra subscribers only. Gemini API access exists solely through an early-access programme that researchers, engineers, and enterprises must apply to. You cannot pick it from the model list and start billing against it.

Its current published figures come from Google’s 12 February 2026 upgrade post: 48.4% on Humanity’s Last Exam without tools, 84.6% on ARC-AGI-2, a Codeforces Elo of 3455, and IMO 2025 gold-medal-level performance. Several secondary write-ups still quote 41.0% and 45.1% — those are the December 2025 launch numbers, not the current ones. No Deep Think refresh has shipped since February 2026.

Omni — the family generates video, it does not just watch it

Gemini Omni was announced at I/O on 19 May 2026 and is a distinct tier, not a Flash variant: text, image, audio, or video in, and video with native audio out, with conversational multi-turn editing. Gemini Omni 1.1 Flash (gemini-omni-1.1-flash) shipped on 27 August 2026, adding first/last-frame interpolation, scene extension in 10-second increments to 40 seconds total, resolution control up to 4K, and a 360p draft mode roughly 60% faster at about a third of the cost — cheap iteration before an upscale. Its launch stage is genuinely ambiguous: Google’s 27 August release note calls it “the GA version” and the deprecations page names it as the replacement for the older gemini-omni-flash-preview endpoint, which shuts down on 30 September 2026 — but the model list on ai.google.dev still files gemini-omni-1.1-flash under Preview. Plan the migration off the preview ID to that deadline regardless, and check the label yourself before you build a change-management argument on it. Consumer access runs through Google Flow for AI Plus, Pro, and Ultra, scene extension in the Gemini app, and free Remix in YouTube Shorts.

Flash Cyber — a frontier variant you cannot simply buy

Alongside 3.8 Flash on 2 September 2026, Google announced Gemini 3.8 Flash Cyber, a cybersecurity-specialised variant for autonomous vulnerability discovery, security research, and automated patching, paired with Google’s CodeMender remediation agent. It is not publicly available. Access runs through the new Fairwind Program, which gives prioritised access to vetted government authorities, critical infrastructure operators, and maintainers of widely used software. Google DeepMind’s programme page sets the terms: applicants must show “a proven track record of ethical operations and research”; participants may grant access only to internal cybersecurity, incident-response, or penetration-testing teams and must enforce phishing-resistant MFA; and partners “may not share, redistribute, or sell access” to the model. Google does not say whether Fairwind participants call it through the ordinary Gemini API, but it does state that zero data retention is available when the model is accessed through the Gemini Enterprise Agent Platform. See Google ships Gemini 3.8 Flash and a gated cybersecurity variant .

Live and speech — the September 2026 voice stack

Google rebuilt its real-time voice line in September 2026, and every piece replaces a preview model that many voice agents still call.

• Gemini 3.8 Live (gemini-3.8-live) and Gemini 3.8 Live Extended Thinking (gemini-3.8-live-extended-thinking) reached general availability on 15 September 2026. Both are audio-to-audio models on the Live API: 3.8 Live is Google’s default for low-latency voice agents, with interleaved reasoning and asynchronous function calling on by default; the Extended Thinking variant adds background reasoning during the conversation. Paid-tier pricing is $0.75 per MTok text / $3.00 audio input and $4.50 text / $12.00 audio output. Google names gemini-3.8-live as the replacement for gemini-3.1-flash-live-preview and the older 2.5 native-audio previews; no shutdown date has been set for those yet. On 24 September Google added Live Avatar — lip-synced, near-real-time video of a visual persona paired with 3.8 Live — available in Gemini Enterprise.
• Gemini 3.8 Flash TTS (gemini-3.8-flash-tts) and Gemini 3.8 Flash-Lite TTS (gemini-3.8-flash-lite-tts) reached GA on 22 September 2026, together with a new Voices endpoint (/v1beta/voices), voice design from text prompts, voice replication with consent verification, and a library of 150+ prebuilt and custom voices. They replace gemini-3.1-flash-tts-preview and the 2.5 TTS previews (no shutdown dates announced).
• Gemini 3.5 Transcribe (gemini-3.5-transcribe) and Gemini 3.5 Transcribe Live (gemini-3.5-transcribe-live) reached GA on 26 August 2026 as dedicated speech-to-text models: 85+ languages, speaker diarization, word-level timestamps, and custom vocabulary biasing up to 1,000 terms; the Live variant streams over WebSockets.

Antigravity Agent — a managed agent with a breaking tool change

The Gemini API’s managed coding agent moved to antigravity-preview-09-2026 on 17 September 2026. The previous antigravity-preview-05-2026 shuts down on 5 October 2026. Integrations that only read final output on a remote sandbox just change the agent string; anything that runs tools locally or parses function_call steps must be rewritten, because the built-in tools were renamed (write_file became write_to_file and replace_file_content, read_file became view_file), parameters switched to PascalCase, file edits now use line-range replacement, and dedicated find_by_name and grep_search tools were added.

Task-specific models on the same API

The tiers above are the general-purpose models, but the Gemini API catalogue is wider. Image generation runs on Nano Banana Pro (gemini-3-pro-image) and Nano Banana 2 (gemini-3.1-flash-image), which have superseded Imagen 4 — deprecated, and already shut down on 17 August 2026, with gemini-3.1-flash-image as Google’s named migration target. Veo 3.1 and Veo 3.1 Lite remain listed for video generation alongside Omni, Lyria 3.5 (lyria-3.5, GA 3 September 2026, full-length songs as 44.1 kHz stereo from text and image input, with Lyria 3 Pro now labelled previous generation) and Lyria RealTime for music, Gemini Embedding 2 for retrieval, and Gemini Robotics ER 2 for embodied reasoning. Several of these carry preview labels; check the models page for the current status of any you intend to ship on.

Gemma — Google’s open-weight family

Gemma is a separate, smaller open-weight family, not a self-hostable Gemini, but it is the answer when Gemini’s hosted-only model is the obstacle. Gemma 4 arrived on 2 April 2026 in E2B, E4B, 26B A4B (mixture-of-experts), and 31B dense sizes, with a unified Gemma 4 12B added on 3 June 2026. Context runs to 128K on E2B and E4B and 256K on the 12B, 26B, and 31B. All sizes handle text and image; E2B, E4B, and 12B also handle audio. The 12B uses a unified, encoder-free architecture and runs locally on a laptop with 16GB of VRAM or unified memory.

The headline is the licence. Gemma 4 is Apache 2.0, replacing the bespoke Gemma Terms of Use that governed Gemma 3 — a real liberalisation, and the thing to check if you evaluated Gemma before April 2026 and ruled it out on licensing. Google also ships specialised variants including EmbeddingGemma, ShieldGemma 2, DiffusionGemma, FunctionGemma, PaliGemma, and RecurrentGemma.

What it costs

Gemini API rates, per million tokens, as published on 25 September 2026:

Watch the Flash price cliff. The $0.75 / $3.75 rate on 3.6, 3.7, and 3.8 Flash is introductory and expires on 31 December 2026. From 1 January 2027 it doubles to $1.50 / $7.50. Any cost model built on today’s Flash price has a step change in it just under four months out. Note also that the older Gemini 3.5 Flash is more expensive than the three newer Flash models, so there is no cost argument for staying on it.

On the consumer side, Google restructured its subscriptions at I/O on 19 May 2026: it added a second AI Ultra level with 5x AI Pro’s usage limits and 20 TB of storage, and cut the top AI Ultra plan’s price. Google’s I/O post gives those as round numbers — a new “$100” tier and the top plan “from $250 to $200” — while the list prices on Google’s own subscriptions page are $99.99 and $199.99/month. Below those sit a free tier, Google AI Plus, and Google AI Pro; independent outlets report Plus at $4.99/month with 400 GB (cut from $7.99 and 200 GB on 8 June 2026) and Pro at $19.99/month with 5 TB, so treat those two numbers as secondary-verified.

What matters architecturally is the gating, and Google’s subscriptions page states it directly: the free tier gets Gemini 3.6 Flash plus limited 3.1 Pro access, Plus is the same with roughly double the limits, Deep Think is Ultra-only, Gemini 3.8 Flash in the app is Pro and Ultra, and Omni video generation reaches down to Plus through Flow credits (200 on Plus, 1,000 on Pro, 10,000–25,000 on Ultra).

How to access it and how it fits

You reach the same underlying models through several surfaces, chosen by who you are and what you are building.

The Gemini API through Google AI Studio suits a solo developer who wants a key and a quick integration. The Gemini Enterprise Agent Platform — the April 2026 rename of Vertex AI, covered in full at Google Vertex AI — suits teams that need regional data controls, identity and access management, and production reliability. Google Cloud now serves its Gemini model documentation under docs.cloud.google.com/gemini-enterprise-agent-platform/, while the consumer app and the developer API keep the Gemini name. Both surfaces call the same model tiers. You choose the surface, not a different model.

Where it sits in a stack: Gemini is the reasoning and generation layer. Your application sends structured or mixed-media input, the model returns text, structured output, or video, and your code handles storage, retrieval, and orchestration around it. It plays the same architectural role that GPT or Claude does in a typical build. See the wider picture in the LLM landscape for 2026 .

Compared to the alternatives

All three families are frontier multimodal models with large context windows. The differences that matter in practice are the access surfaces, the cloud you are already on, and the tooling around each.

Treat this table as a starting point for a shortlist, not a verdict. Model rankings shift with each release — Google alone shipped three Flash models between July and September 2026 — so benchmark the current tiers on your own workload before committing. Compare Claude and the Azure-hosted GPT option in Azure OpenAI alongside Gemini.

When not to use it

Gemini is not always the right call.

• You are standardised on AWS with no Google Cloud footprint. If your data, identity, and networking all live in AWS, a model served through Amazon Bedrock or Amazon Nova keeps traffic and governance in one place.
• You need a fully self-hosted or open-weights model. The hosted Gemini models are proprietary. Google’s own answer is Gemma 4 — Apache 2.0 since 2 April 2026, five sizes from E2B to a 31B dense model, free to download and run — but it is a smaller, separate family, not a self-hosted Gemini. If you need frontier-tier capability on your own hardware, Gemini is not the family.
• You are starting a new project on Gemini 2.5. Since 18 September 2026 the 2.5 models are available only to prior users; start on 3.8 Flash or 3.5 Flash-Lite instead.
• Your compliance process bars preview endpoints. The Pro tier is only available as gemini-3.1-pro-preview. If preview status is a blocker, your Google options are the GA Flash line or nothing at that capability level.
• You need Deep Think or Flash Cyber on ordinary commercial terms. Deep Think is an AI Ultra subscription feature with API access by application only; Flash Cyber is restricted to Fairwind Program participants. Neither is something you can procure by adding a credit card.
• Your task is narrow and small. A frontier multimodal model is overkill for simple classification or extraction that a small specialised model — or a Gemma 4 E2B running locally — handles at a fraction of the cost.
• You cannot send data to a third-party API. If regulation forbids sending inputs off-premises, a hosted API of any vendor is a poor fit.

Match the tier to the task even when Gemini is the right family. Flash-Lite for high volume, Flash for most agentic and coding work, Pro for the hardest reasoning you can actually reach through the API. Paying Pro rates for a Flash-Lite job wastes money — and the Flash line is now capable enough that the case for reaching past it is narrower than it was six months ago.

Further reading

• Foundation models : what a general-purpose pretrained model is and why tiers exist.
• What is an LLM : the language model concepts behind every tier of Gemini.
• Google Vertex AI : the enterprise surface and what the April 2026 rename to Gemini Enterprise Agent Platform changed.
• Google ships Gemini 3.8 Flash and a gated cybersecurity variant : the 2 September 2026 release and the Fairwind Program.
• Google I/O 2026 : where Gemini Omni, agentic Search, and the subscription restructure were announced.
• Claude by Anthropic : the closest comparable frontier family, useful for a side-by-side trial.
• The LLM landscape in 2026 : how Gemini, GPT, and Claude compare across the market.
• Gemini API models : the current, authoritative list of model versions and identifiers.

Sources

• Gemini API models: https://ai.google.dev/gemini-api/docs/models — current model list, model IDs, and stable versus preview status.
• Gemini API pricing (fetched 25 September 2026): https://ai.google.dev/gemini-api/docs/pricing — per-MTok rates, 3.8 Live audio and text rates, and the introductory Flash window expiring 31 December 2026.
• Gemini API deprecations (fetched 25 September 2026): https://ai.google.dev/gemini-api/docs/deprecations — the gemini-omni-flash-preview shutdown on 30 September 2026, antigravity-preview-05-2026 on 5 October 2026, gemini-2.5-flash-image on 2 October 2026 and gemini-3.1-flash-lite on 7 May 2027, the Gemini 2.5 access note, and the Live and TTS replacement mappings.
• Gemini API release notes (fetched 25 September 2026): https://ai.google.dev/gemini-api/docs/changelog — dated entries for 3.8/3.7/3.6 Flash, 3.5 and 3.1 Flash-Lite, Omni 1.1 Flash, 3.5 Transcribe (26 August), Lyria 3.5 (3 September), 3.8 Live (15 September), Antigravity 09-2026 (17 September), the Gemini 2.5 access restriction (18 September), and 3.8 Flash TTS / Flash-Lite TTS (22 September).
• Google, “Introducing Gemini 3.8 Live with Live Avatar”, 24 September 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-with-live-avatar/
• Gemini 3.8 Flash and 3.8 Flash Cyber announcement, 2 September 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/
• The Fairwind Program: https://blog.google/innovation-and-ai/technology/safety-security/fairwind-program/ — the announcement and CodeMender pairing.
• Google DeepMind, Fairwind Program page: https://deepmind.google/fairwind-program/ — the eligibility groups, the “proven track record of ethical operations and research” vetting standard, and participant obligations.
• Gemini 3.1 Pro announcement, 19 February 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/
• Gemini 3 Deep Think upgrade, 12 February 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think/
• Google DeepMind, Gemini Flash: https://deepmind.google/models/gemini/flash/ — general availability and the 1M/64K context figures.
• Google DeepMind, Gemini Pro: https://deepmind.google/models/gemini/pro/ — 3.1 Pro as the current Pro tier and the “3.5 Pro coming soon” note.
• Google AI subscriptions at I/O 2026: https://blog.google/products-and-platforms/products/google-one/google-ai-subscriptions/ — the restructure and the rounded Ultra price points.
• Google AI plans and pricing: https://gemini.google/subscriptions/ — list prices, per-tier model access, and Flow credit allowances.
• Gemma 4 model card: https://ai.google.dev/gemma/docs/core/model_card_4 — sizes, parameter counts, context windows, modalities, and the Apache 2.0 licence.
• Gemma 4 12B developer guide, 3 June 2026: https://developers.googleblog.com/gemma-4-12b-the-developer-guide/
• Google Cloud model docs under the renamed platform: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-1-pro