Skip to content

Google AI models: Gemini and Gemma explained

  • How Google’s portfolio splits into hosted Gemini models and open-weight Gemma models.
  • The Gemini lineup as of November 2026: model IDs, context limits, thinking levels and where each model fits.
  • The two ways to reach Gemini (the Gemini API and Google Cloud) and why the choice matters.
  • How Gemma sizes and licenses work, including what the “E” and “A” in the names mean.
  • What goes wrong in production: aggressive deprecation, preview models, introductory prices and silent rerouting.

Google offers two families. Gemini models are proprietary and hosted: you call them through the Gemini API (with Google AI Studio as the console) or through Google Cloud. Gemma models are open-weight: you download them and run them on your own hardware or any cloud.

As of November 2026 the current Gemini generation is Gemini 4, with three models named after elements: Hydrogen, Argon and Carbon, in rising order of capability and price. Gemini 3.8 Flash remains the current Flash model from the previous numbering. The current open family is Gemma 4.

Google changes this lineup more often than most providers and retires models on short notice, so this page separates the durable parts (how the families work, how to choose) from a dated snapshot.

Family What it is Where it runs License
Gemini 4 (Hydrogen, Argon, Carbon) Current hosted generation Gemini API, Google AI Studio, Google Cloud Proprietary service
Gemini 3.x Flash and Flash-Lite Previous numbering, still the fast and low-cost line Gemini API, Google Cloud Proprietary service
Media and specialist models Image, video, music, speech, live audio, embeddings Gemini API, Google Cloud Proprietary service
Gemma 4 Open-weight general models Your hardware, or any host Apache 2.0
Gemma 3 and derivatives (MedGemma, ShieldGemma 2) Earlier open-weight and specialist models Your hardware Gemma Terms of Use or model-specific terms

Three ideas explain most of Google’s naming:

  • Tier words describe size and speed. Pro (largest), Flash (fast and cheaper), Flash-Lite (cheapest, highest volume). Gemini 4 replaced these words with element names, but the same ladder exists.
  • Stable versus preview. A stable model ID has no suffix and usually does not change. Preview IDs carry a -preview suffix, can change, and are shut down with short notice.
  • Thinking levels. Recent Gemini models reason before answering, and you choose how much with a thinking level. Thinking tokens are billed as output tokens.

Sources: Google’s Gemini API models page, Google Cloud model documentation and the official Gemini cookbook. Recheck before relying on any row.

Model Model ID Released Positioning Thinking levels
Gemini 4 Carbon gemini-4-carbon Announced Oct 28, 2026; generally available Nov 2026 Top tier: strongest reasoning, hardest multi-step problems Low, medium, high (as reported at launch)
Gemini 4 Argon gemini-4-argon Oct 13, 2026 Frontier tier: most demanding engineering and knowledge work Low, medium, high (default medium)
Gemini 4 Hydrogen gemini-4-hydrogen Nov 12, 2026 Balanced tier: Google’s suggested starting point for everyday coding, multimodal work and agents Four levels up to xhigh (default high)
Gemini 3.8 Flash gemini-3.8-flash 2026 (see release notes) Balanced tier: the most capable Flash model Low, medium, high (default medium)
Gemini 3.5 Flash-Lite gemini-3.5-flash-lite Generally available Jul 2026 Fast tier: lowest-cost, high-volume and subagent work See the model page

Shared characteristics, per Google’s documentation:

  • Context: about 1 million tokens (1,048,576). On Gemini 4 this is a single limit covering input and output together, and thinking counts as output.
  • Output: on Gemini 4, up to 65,536 tokens in one response segment when streaming (Google Cloud documents 32,768 for non-streaming requests), with a “long decode” continuation feature for longer outputs. Gemini 3.8 Flash lists a 64k maximum output.
  • Input types: text, images, video, audio and PDF. Output from these models is text; images, video and audio come from separate models.
  • Tools: function calling, structured output, code execution and grounding with Google Search.
  • Knowledge cutoff: March 2026 for Gemini 4 Argon, per its model page; check each model’s page for the others.

Carbon and Argon also have allowlist-only variants for approved defensive cybersecurity work.

Prices quoted by Google in November 2026

Paid-tier prices per million tokens, as stated in Google’s launch posts and pricing pages on the dates checked. Several are introductory and change on a fixed date, so read the Gemini API pricing page before budgeting.

Model Input / output Notes
Gemini 4 Carbon $2 / $10 introductory $4 / $20 after November 30, 2026; higher long-context rates when input exceeds 200K tokens on Google Cloud (as reported at launch)
Gemini 4 Argon $2 / $10 introductory $4 / $20 after November 30, 2026
Gemini 4 Hydrogen $1.50 / $7.50 Stated as flat across the context window; also offered on the free tier
Gemini 3.8 Flash $0.75 / $3.75 introductory (as reported at launch) $1.50 / $7.50 from January 1, 2027 (as reported at launch)

Google ships separate models for other modalities, and these change name and status even faster than the text models. As of November 2026 the Gemini API documentation lists an image family (marketed as Nano Banana), a video and audio generation model family (Gemini Omni Flash, which replaced the Veo 3.1 preview models shut down in October 2026), Lyria for music, Live models for real-time voice, text-to-speech models, and Gemini Embedding models for retrieval. Take the exact IDs from the models page at the time you build.

Gemini API (Google AI Studio) Google Cloud
Audience Individual developers, startups, prototypes Enterprises already on Google Cloud
Authentication API key Google Cloud IAM and projects
Free tier Yes, for selected models, with rate limits Billed through a Cloud project; see Google Cloud pricing
Controls Simpler; fewer enterprise controls Regions, VPC controls, provisioned throughput, enterprise data terms
Lifecycle Its own deprecation schedule Its own, different, retirement dates

Google Cloud’s AI platform was renamed in April 2026: Vertex AI is now Gemini Enterprise Agent Platform, and Google’s older Vertex AI documentation is no longer updated. You will still see “Vertex AI” in SDK parameters, older articles and many tools.

The same google-genai SDK targets both. This fragment requires an API key and was not executed; the call shape is taken from the SDK’s README (version 3.3.0 on PyPI).

# Fragment: requires GEMINI_API_KEY in the environment; not executed
# pip install google-genai
from google import genai
client = genai.Client() # reads GEMINI_API_KEY (or GOOGLE_API_KEY)
response = client.models.generate_content(
model="gemini-3.8-flash",
contents="Summarize in one sentence: order A1001 arrived damaged and the customer wants a replacement.",
)
print(response.text)

One practical trap: data-handling terms differ between the free tier and paid usage. Read the terms for the tier you are on before sending customer data.

Gemma models are built from the same research as Gemini but released as weights you run yourself, through tools such as Ollama, llama.cpp, vLLM or Hugging Face Transformers. See Running LLMs locally for the runtime side.

Released in April 2026 under the Apache 2.0 license. Sizes on the current model card:

Size What the name means Context Input
E2B About 2.3B “effective” parameters (about 5.1B including embeddings) 128K Text, image, audio
E4B About 4.5B effective parameters (about 8B including embeddings) 128K Text, image, audio
12B Dense model, added June 2026 256K Text, image, audio
26B A4B Mixture of experts: about 26B total parameters, about 4B active per token 256K Text, image
31B Dense model 256K Text, image
  • “E” means effective. The small models use per-layer embeddings, large lookup tables that add to the stored size without adding to the compute per token. Plan memory from the total size, not the effective count.
  • “A” means active. A mixture-of-experts model runs only part of its parameters for each token, so it computes like a small model, but all 26B parameters must still be loaded into memory.

Gemma 3 (2025, in sizes up to 27B) and the mobile-oriented Gemma 3n remain available. Unlike Gemma 4, they are distributed under the Gemma Terms of Use, which include a prohibited-use policy, not under Apache 2.0.

  • MedGemma is a Gemma 3 based family for medical text and image understanding. Its use is governed by the Health AI Developer Foundations terms. It is a starting point for developers, and any clinical use needs validation and the relevant regulatory process.
  • ShieldGemma 2 is a 4B image safety classifier built on Gemma 3, for screening images against safety policies.

Always read the license on the specific model card. “Open” in this family covers two different legal arrangements.

Need Start with Why
General assistant, everyday coding, tool-using agents Gemini 4 Hydrogen Google’s suggested starting point; thinking up to xhigh
Cost-sensitive volume before January 2027 Gemini 3.8 Flash Half Hydrogen’s price until its introductory rate ends (as reported)
Complex engineering, analysis, multi-step agents Gemini 4 Argon Google’s recommendation for most work that needs more reasoning than Hydrogen
The hardest reasoning where cost is secondary Gemini 4 Carbon Reserved by Google for tasks that test the limits of a model
Very high volume classification, extraction, subagents Gemini 3.5 Flash-Lite Lowest cost tier
Long video, audio or very large document input Any current Gemini model Native multimodal input with a 1M-token context
Data must stay on your hardware; fine-tuning Gemma 4 (size to fit your memory) Open weights under Apache 2.0
Phones and edge devices Gemma 4 E2B or E4B, or Gemma 3n Built for small memory footprints
Image safety screening ShieldGemma 2 Purpose-built classifier

Hydrogen or 3.8 Flash? Both sit in the balanced tier. Hydrogen is the newer generation, Google’s suggested starting point, and has a fourth thinking level (xhigh) for harder agent steps. 3.8 Flash costs half as much until January 1, 2027, after which the reported prices are the same ($1.50 / $7.50). Prefer 3.8 Flash for high-volume work where your evaluation shows it passes and the saving matters before that date; prefer Hydrogen for new builds, agents and anything that will still run in 2027, so you do not migrate twice.

Then test. Run a graded set of your own tasks against two neighboring tiers and two thinking levels, and choose the cheapest combination that passes.

Worked example: a support assistant on Gemini

Section titled “Worked example: a support assistant on Gemini”

An invented shop, OrdersPortal, builds a support assistant. The design uses three tiers instead of one:

  1. Routing and intent classification runs on Flash-Lite. The prompts are short and the volume is high.
  2. Answering from policy documents runs on Gemini 3.8 Flash at a low or medium thinking level, with the policy text supplied in the prompt and Search grounding switched off so answers come only from company content.
  3. Disputed refunds with long histories and photo evidence escalate to Gemini 4 Argon, which takes the images and the full conversation in one request.

The bill for steps 2 and 3, at the prices in the collapsible above. Volumes and token counts are invented for illustration; the routing step on Flash-Lite is small and left out.

Step Monthly volume Tokens per request (input / output incl. thinking) Monthly tokens Until Nov 30, 2026 From Dec 1, 2026 From Jan 1, 2027
2. Policy answers on 3.8 Flash 300,000 8,000 / 500 2,400M / 150M $2,363 $2,363 $4,725
3. Disputes on Argon 20,000 60,000 / 3,000 1,200M / 60M $3,000 $6,000 $6,000
Total $5,363 $8,363 $10,725

The same design costs twice as much by January without a line of code changing. Two price steps hit two different steps of the pipeline on two different dates, so the budget needs each model’s price schedule, not a single rate. The levers: move step 2 to Hydrogen now if it passes the same evaluation (the January price is the same, and you avoid a later migration), lower Argon’s thinking level for disputes that do not need it, and trim the history sent with each dispute.

Because thinking tokens are billed as output, the team sets the thinking level per step and caps output length in the prompt. They store model IDs in configuration, since each of these models will be replaced within a year.

Choice Gain Cost
Gemini (hosted) Native video and audio input, 1M-token context, Search grounding, a free tier to start Fast-moving lineup, short deprecation windows, introductory prices that step up
Gemini API Quick start with an API key Fewer enterprise controls than Google Cloud
Google Cloud IAM, regions, enterprise terms More setup; separate retirement dates
Gemma (self-hosted) Data control, no per-token fee, fine-tuning You run the hardware; below hosted frontier quality on hard tasks
Higher thinking level Better results on hard problems More output tokens and latency
Another provider Different strengths and pricing Prompts and tool behavior need re-testing

For comparison across providers, see LLM model selection, Anthropic models, OpenAI models and Amazon models.

  • Short model lifetimes. Gemini 2.5 models reached retirement on Google Cloud in October 2026, and Google’s pages gave differing dates. Subscribe to the deprecations page for the surface you use, and budget migration time every few months.
  • Silent rerouting. Google’s changelog records deprecated IDs being routed to a newer model (for example gemini-3.7-flash to gemini-3.8-flash in October 2026). Behavior and price can change under an unchanged ID. Log the model version returned in responses and re-run evaluations after each announcement.
  • Preview in production. Preview models have tighter rate limits and can be withdrawn at short notice. Several of Google’s strongest models stayed in preview for months.
  • Introductory pricing. A budget made at launch prices can double on the published step-up date.
  • Two schedules. The Gemini API and Google Cloud deprecate on different dates and sometimes expose different features. A fix on one does not carry to the other.
  • Feature gaps between generations. Google’s documentation notes that explicit context caching is not supported on Gemini 4 Argon and later, which rely on implicit caching. Check each feature you depend on when moving up a generation.
  • Thinking cost. Thinking tokens count as output and against the shared context limit on Gemini 4. A high level on a simple task wastes money and time.
  • Naming confusion. Vertex AI versus Gemini Enterprise Agent Platform, tier words versus element names, and stable versus preview IDs cause real configuration errors. Copy IDs from the official page, not from memory or blog posts.
  • Gemma license mix-ups. Assuming Apache 2.0 for a Gemma 3 based model, or for MedGemma, can breach its terms.
  • Gemini, when your organization cannot absorb frequent model migrations, or when data must not leave your infrastructure and Google Cloud’s enterprise terms are not acceptable.
  • A Gemini preview model, for anything with an uptime commitment.
  • Carbon, for routine tasks: Hydrogen or Flash will be faster and far cheaper.
  • Gemma, when the task needs frontier-level reasoning, or when nobody will operate the serving infrastructure.
  • MedGemma, as a clinical decision tool without validation and regulatory work.

1. What is the difference between Gemini and Gemma? Gemini is Google’s proprietary, hosted model family, reached through the Gemini API or Google Cloud and billed per token. Gemma is the open-weight family you download and run yourself. Gemini offers the strongest capability and the largest context; Gemma offers control over data, cost and customization.

2. How do the Gemini tiers relate to each other? They form a ladder of capability against speed and price. In the 3.x numbering the rungs are Flash-Lite, Flash and Pro. In Gemini 4 they are Hydrogen, Argon and Carbon. You pick the lowest rung that passes your evaluation and adjust the thinking level before moving up.

3. Gemini API or Google Cloud? The Gemini API is the fastest way to start: an API key, a free tier and a simple console. Google Cloud is for production in an enterprise: IAM, regional control, networking controls and enterprise data terms. The same SDK reaches both, but lifecycle dates and some features differ, so choose deliberately and do not assume parity.

4. What do “E4B” and “26B A4B” mean in Gemma 4 names? “E” is effective parameters: E4B computes like a model of about 4.5B parameters but stores about 8B because of per-layer embedding tables. “A” is active parameters in a mixture-of-experts model: 26B A4B activates about 4B parameters per token, giving small-model speed, but needs memory for all 26B.

5. Senior scenario: a team wants to use Gemini’s 1M-token context to “just send the whole knowledge base” instead of building retrieval. How do you respond? It can be right for a bounded, slowly changing corpus that fits, and it removes retrieval errors. But every request pays for the whole context, latency grows with input size, long-context rates may apply above a threshold, and accuracy on details can fall as context grows. I would measure both designs on real questions: cost per answer, latency and answer quality. Context caching changes the economics for repeated prefixes, but its availability differs by model generation. For a large or frequently updated corpus, retrieval with a smaller context usually wins on cost and freshness.

  • Gemini is hosted and proprietary; Gemma is open-weight and self-hosted.
  • As of November 2026: Gemini 4 Hydrogen, Argon and Carbon, plus Gemini 3.8 Flash and 3.5 Flash-Lite, all with about a 1M-token context.
  • Choose the lowest tier and thinking level that pass your own evaluation.
  • The Gemini API and Google Cloud (Gemini Enterprise Agent Platform, formerly Vertex AI) are separate surfaces with separate lifecycles.
  • Expect frequent deprecations, rerouted IDs and introductory prices; keep IDs in configuration and evaluations ready.
  • Gemma 4 is Apache 2.0; Gemma 3, MedGemma and ShieldGemma have their own terms.

Checked on November 13, 2026. ai.google.dev is not directly reachable from the verification environment, so Google pages were checked through search results on official Google domains; figures marked “as reported at launch” come from launch coverage and should be confirmed on the pricing page.