Gemini 4 is Google's newest generation of Gemini models, made up of Gemini 4 Argon, Gemini 4 Carbon and Gemini 4 Hydrogen, released between October and November 2026, and it changes how the API behaves compared with Gemini 3.1 Pro, the preview model it succeeds. This page sets the three Gemini 4 models beside gemini-3.1-pro-preview: model IDs, release dates, launch stage, context and output limits, thinking controls, prices, regions, published benchmarks and lifecycle dates. It is written for developers migrating code off Gemini 3.1 Pro, architects planning capacity and buyers estimating cost. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from Google's own documentation, pricing page, changelog and announcements, cited under Standards and references.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Google's public documentation as of November 2026. Model details change, so confirm against Google's current documentation.
Key takeaways
- Gemini 4 is a family, not one model. Google lists three generally available Gemini 4 models: Argon (GA 13 October 2026), Carbon (GA 10 November 2026) and Hydrogen (GA 12 November 2026). Gemini 3.1 Pro is still a public preview, released on 19 February 2026.
- Input and output now share one 1,048,576-token limit. Gemini 3.1 Pro has a 1,048,576-token input limit and a separate 65,536-token output limit; Gemini 4 models can extend output towards the full window with long decode.
- Several 3.1 Pro parameters break on Gemini 4. Google's Argon guide states that thinking_budget, MINIMAL thinking, frequency_penalty, presence_penalty and candidate_count return HTTP 400, and that temperature, top_p and top_k are ignored.
- Pricing changed shape. Gemini 3.1 Pro charges $2 and $12 per million input and output tokens up to 200K prompt tokens and more above it; Argon and Carbon charge a flat $2 and $10 until 30 November 2026, then $4 and $20. Hydrogen lists $1.50 and $7.50.
- No retirement date is published for 3.1 Pro. The Gemini API changelog lists no deprecation of gemini-3.1-pro-preview as of 15 November 2026; Argon's Google Cloud retirement date is 13 October 2027 or later.
What is Gemini 4?
Gemini 4 is the generation of proprietary Gemini models that Google began shipping in autumn 2026. Google's Argon developer guide calls Argon the first model in the Gemini 4 family. Google Cloud describes Carbon as its frontier model for deep work on complex tasks, and Google's Carbon announcement of 10 November 2026 states that "Argon remains the ideal choice for the majority of software engineering and knowledge work", while Carbon uses more tokens to reason and verify on harder tasks. Google Cloud describes Hydrogen as "built for speed and efficiency across everyday coding and agentic workflows". All three are thinking models that accept text, image, audio and video input and return text.
Gemini 3.1 Pro is the last Pro-named model in Google's Gemini 3 series. The Gemini API changelog entry for 19 February 2026 describes it as the latest iteration in the Gemini 3 series, and on 9 March 2026 the older gemini-3-pro-preview ID was shut down and pointed to gemini-3.1-pro-preview. Google has not released a Gemini 4 model under the Pro name; Google's Argon guide describes migration into Argon from gemini-3.1-pro-preview. For the full Google line-up see Google Gemini Models, and for differences inside the new family see Gemini 4 Carbon vs Argon vs Hydrogen.
Every model at a glance
As of 2026-11-15. Sources: Google Cloud model pages for each model, Google AI Studio model pages and the Gemini API changelog.
| Model | API ID | Released | Status |
|---|---|---|---|
| Gemini 4 Hydrogen | gemini-4-hydrogen | 2026-11-12 | GA (Google Cloud); Stable (AI Studio) |
| Gemini 4 Carbon | gemini-4-carbon; gemini-4-carbon-cyber-permissive (allowlist) | 2026-11-10 | GA |
| Gemini 4 Argon | gemini-4-argon; gemini-4-argon-cyber-permissive (allowlist) | 2026-10-13 | GA |
| Gemini 3.1 Pro | gemini-3.1-pro-preview; gemini-3.1-pro-preview-customtools | 2026-02-19 (customtools 2026-02-23) | Public preview |
Open weights: none for any of the four. Licence: proprietary, under Google's Gemini API and Google Cloud terms. Parameters: not stated by Google for any of the four on the pages checked. Google publishes no dated snapshot suffix for any Gemini 4 model.
When was each model released?
As of 2026-11-15. Sources: Gemini API changelog, Google Cloud model pages and Google's announcements. Oldest first.
| Date | Event | Model IDs |
|---|---|---|
| 2026-02-19 | Gemini 3.1 Pro Preview released on the Gemini API and Google Cloud | gemini-3.1-pro-preview |
| 2026-02-23 | Customtools endpoint released on Google Cloud, tuned for mixed bash and custom tool workflows | gemini-3.1-pro-preview-customtools |
| 2026-03-09 | Gemini 3 Pro Preview shut down; its ID points to gemini-3.1-pro-preview | gemini-3-pro-preview |
| 2026-10-13 | Gemini 4 Argon generally available on the Gemini API and Google Cloud | gemini-4-argon |
| 2026-11-10 | Gemini 4 Carbon generally available; announcement "Gemini 4 Carbon is now available" | gemini-4-carbon |
| 2026-11-12 | Gemini 4 Hydrogen generally available on Google Cloud | gemini-4-hydrogen |
| 2026-12-01 | Announced price rise for Argon and Carbon takes effect | gemini-4-argon, gemini-4-carbon |
The gap between the two generations is under eight months. Google's Carbon announcement states that a cyber variant of Carbon was offered to defenders through the Fairwind Program before its general release. For release dates across all providers see the LLM Release Tracker.
How much does each model cost?
As of 2026-11-15. USD per million tokens, Gemini API paid tier, from the Google AI Studio pricing page. Output prices include thinking tokens.
| Model | Input | Output | Cached input | Batch and Flex (input / output) | Priority (input / output) |
|---|---|---|---|---|---|
| Gemini 4 Hydrogen | $1.50 | $7.50 | $0.075 | $0.75 / $3.75 | $2.70 / $13.50 |
| Gemini 4 Carbon | $2.00, then $4.00 from 1 Dec 2026 | $10.00, then $20.00 | $0.10, then $0.20 | $1.00 / $5.00, then $2.00 / $10.00 | $3.60 / $18.00, then $7.20 / $36.00 |
| Gemini 4 Argon | $2.00, then $4.00 from 1 Dec 2026 | $10.00, then $20.00 | $0.10, then $0.20 | $1.00 / $5.00, then $2.00 / $10.00 | $3.60 / $18.00, then $7.20 / $36.00 |
| Gemini 3.1 Pro | $2.00 (up to 200K); $4.00 (over 200K) | $12.00; $18.00 | $0.20; $0.40, plus $4.50 per 1M tokens per hour storage | $1.00 / $6.00; $2.00 / $9.00 | $3.60 / $21.60; $7.20 / $32.40 |
- Long prompts: Gemini 3.1 Pro prices step up for prompts over 200K tokens. The pricing page lists no separate long-prompt tier for Argon, Carbon or Hydrogen.
- Caching: Gemini 3.1 Pro supports explicit context caching with a storage charge. Argon, Carbon and Hydrogen support implicit caching only, so there is no storage line.
- Free tier: Argon, Carbon and Gemini 3.1 Pro are not on the Free tier. The pricing page lists Hydrogen as free of charge on the Free tier for Standard and Priority.
- Contributor tier: an invite-only preview tier priced at Batch rates for Argon and Hydrogen (none listed for 3.1 Pro), in which content is used to improve Google's products.
- Grounding: Google Search and Google Maps grounding each include 5,000 free requests a month shared across Gemini 3 and newer models, then $14 per 1,000.
Illustrative arithmetic: a request of 100,000 input and 10,000 output tokens costs $0.32 on Gemini 3.1 Pro at Standard rates, $0.30 on Argon until 30 November 2026 and $0.60 from 1 December 2026, and $0.225 on Hydrogen, before thinking-token differences. Per-task cost also depends on how many thinking tokens each model spends. For the mechanics see Model Pricing Explained.
What changed between models?
Google's Gemini 4 Argon developer guide lists the API changes for code moving from gemini-3.1-pro-preview. The Carbon and Hydrogen model pages show the same thinking and sampling rules.
| Area | Gemini 3.1 Pro | Gemini 4 (Argon, Carbon, Hydrogen) |
|---|---|---|
| Token limits | Separate limits: 1,048,576 input, 65,536 output | One unified 1,048,576-token limit; max_output_tokens above it returns HTTP 400 |
| Default output per request | Up to 65,536 | 65,536 streamed, 32,768 non-streamed per foreground segment (Argon, Carbon); Hydrogen 65,536 by default |
| Long decode | Not listed on the model pages | Output up to 1,048,576 minus input on the Interactions API or v1beta1 GenerateContent |
| thinking_budget | Accepted | Returns HTTP 400 |
| thinking_level MINIMAL | Available | Returns HTTP 400 |
| temperature, top_p, top_k | Configurable (temperature 0.0 to 2.0, default 1.0) | Ignored; fixed by the backend |
| frequency_penalty, presence_penalty, candidate_count | candidateCount 1 to 8 on Google Cloud | Return HTTP 400; remove them |
| Prefilled model turns | Not listed as a restriction | Not supported; history cannot end with a model turn |
| Context caching | Implicit and explicit | Implicit only, for shared prefixes of 4,096 tokens or more; remove cached_content calls |
| Computer use | Not supported | Supported on the standard models; not on the cyber permissive variants |
| RAG Engine (Google Cloud) | Supported | Not supported |
Google's migration steps, in order: change the model ID; upgrade the Google Gen AI SDK and set max_output_tokens above 65,536 where longer output is needed; replace thinking_budget and MINIMAL with thinking_level; remove explicit caching and put static content of 4,096 tokens or more first; remove sampling and penalty parameters; and remove prefilled model turns. The guide also states that every processed token is billed, including thinking tokens and internal tool instructions used with grounding.
What workloads does Google position each model for?
| Model | Positioning in Google's words or documentation | Source |
|---|---|---|
| Gemini 4 Hydrogen | "Built for speed and efficiency across everyday coding and agentic workflows" | Google Cloud model page |
| Gemini 4 Carbon | "Our strongest model with frontier performance in cybersecurity capabilities"; for harder tasks such as complex multi-step coding and open-ended research | Carbon announcement, 10 Nov 2026 |
| Gemini 4 Argon | "Most capable model for complex coding, agentic, math, and science tasks" (changelog, 13 Oct 2026); "the ideal choice for the majority of software engineering and knowledge work" (Carbon announcement) | Gemini API changelog; Carbon announcement |
| Gemini 3.1 Pro | "Our most advanced reasoning model" (Google Cloud page description); latest iteration in the Gemini 3 series (changelog, 19 Feb 2026) | Google Cloud model page; Gemini API changelog |
Positioning statements are Google's own and are dated to when each page was written; the 3.1 Pro description predates the Gemini 4 releases.
What is the context window of each model?
As of 2026-11-15. Sources: Google Cloud and AI Studio model pages.
| Model | Context window | Max output | Thinking levels (default) | Knowledge cutoff |
|---|---|---|---|---|
| Gemini 4 Hydrogen | 1,048,576, shared with output | 65,536 by default; up to 1,048,576 with long decode | LOW, MEDIUM, HIGH, XHIGH (HIGH) | March 2026 |
| Gemini 4 Carbon | 1,048,576, shared with output | 65,536 streamed, 32,768 non-streamed per segment; up to 1,048,576 with long decode | LOW, MEDIUM, HIGH (MEDIUM) | March 2026 |
| Gemini 4 Argon | 1,048,576, shared with output | 65,536 streamed, 32,768 non-streamed per segment; up to 1,048,576 with long decode | LOW, MEDIUM, HIGH (MEDIUM) | Not stated by Google on the pages checked |
| Gemini 3.1 Pro | 1,048,576 input | 65,536 | Thinking supported; Google Cloud notes MEDIUM was added as a thinking_level option | Not stated by Google on the pages checked |
The window size is the same across both generations; what changed is that Gemini 4 counts input, thinking and output against one budget. A request that fills most of the window with input leaves less room for output. See Context Windows Explained.
Capabilities and tools compared
| Capability | Gemini 4 Hydrogen | Gemini 4 Carbon | Gemini 4 Argon | Gemini 3.1 Pro |
|---|---|---|---|---|
| Input modalities | Text, image, audio, video, PDF | Text, image, audio, video | Text, image, audio, video | Text, image, audio, video, PDF |
| Output | Text | Text | Text | Text |
| Function calling and structured output | Supported | Supported | Supported | Supported |
| Code execution and URL context | Supported | Supported | Supported | Supported |
| Grounding | Google Search, Google Maps | Supported | Supported | Google Search, Google Maps; on Google Cloud also Parallel and Exa web search |
| Computer use | Supported | Supported | Supported | Use a Gemini 4 model for computer use |
| Agentic video understanding | Supported (static by default) | Supported | Supported (static by default) | Use a Gemini 4 model for agentic video |
| APIs | GenerateContent, Interactions, Count Tokens, Chat Completions | GenerateContent, Interactions, Count Tokens, Chat Completions | GenerateContent, Interactions, Count Tokens, Chat Completions | GenerateContent, Interactions, Count Tokens, Chat Completions |
| Tuning and Live API | Not offered | Not offered | Not offered | Not offered |
Modality lists follow each model's own page, which is why PDF appears on some rows and not others. Computer use brings the agent risks covered in the Computer-Use and Browser Agents guide.
Where is each model available?
| Model | Gemini API and AI Studio | Google Cloud regions | Consumer apps |
|---|---|---|---|
| Gemini 4 Hydrogen | Yes, including the Free tier | global, us, eu | Not stated on the model pages checked |
| Gemini 4 Carbon | Paid tier; also Android Studio | global, us, eu; verified enterprise customers in select regions | Gemini app, AI Mode and Antigravity for Google AI Ultra subscribers |
| Gemini 4 Argon | Paid tier | global, us, eu, europe-west2, asia-southeast1 (Standard and Priority PayGo in global, us, eu) | Not stated on the model pages checked |
| Gemini 3.1 Pro | Paid tier | global | Not stated on the model pages checked |
Restricted variants: gemini-4-argon-cyber-permissive and gemini-4-carbon-cyber-permissive are allowlist-only through Google's Fairwind Program for defensive security work, default to HIGH thinking and do not support computer use. Google states access to earlier Gemini cyber models does not carry over. For key and project controls see the Google Gemini API and AI Studio Security Guide.
Model cards and system cards
- Gemini 4 Hydrogen: the AI Studio model page links a model card.
- Gemini 4 Carbon: Google's announcement links a Gemini 4 Carbon model card and a separate Gemini 4 Carbon Safety and Security report. The announcement reports that external testing by Trajectory Labs found no new universal jailbreaks across 132 hours and more than 13,000 requests, and that prompt injection robustness "remains an area for continued improvement".
- Gemini 4 Argon: the Carbon model card refers readers to the Argon model card for architecture and training details. See the Gemini 4 Argon profile for its documentation set.
- Gemini 3.1 Pro: no model card is linked from the Gemini API or Google Cloud model pages checked.
For what these documents contain, see Model Cards and System Cards Explained.
Benchmarks
Provider-reported only: all scores are published by Google on its DeepMind Gemini models page, as shown on 2026-11-15; the page does not date its tables, and Hydrogen appears in a different table from Argon and Carbon. Google publishes no Gemini 3.1 Pro scores on the pages checked, so no generation-on-generation row can be built from Google's own tables.
| Benchmark (as named by Google) | Gemini 4 Hydrogen | Gemini 4 Carbon | Gemini 4 Argon | Gemini 3.1 Pro |
|---|---|---|---|---|
| Agents' Last Exam v1.1 | 71.6% | 77.4% | 69.0% | Not stated by Google |
| OSWorld-2.1 (offline subset) | 77.6% | 81.5% | 74.7% | Not stated by Google |
| Chartography | 66.6% | 72.5% | 71.6% | Not stated by Google |
| Vending-Bench 2 (bank balance) | $14,549 | $16,484 | $13,718 | Not stated by Google |
Only rows where Google lists all three Gemini 4 models are shown. The DeepMind page also quotes a customer, Box, reporting a 14% accuracy improvement for Argon over Gemini 3.1 Pro on its own evaluation; that is a customer statement, not a Google benchmark. Independent leaderboard figures are not included. See How to Read LLM Benchmarks.
Deprecations and retirements
| Model or item | Details | Source |
|---|---|---|
| gemini-4-hydrogen | GA; retirement date not stated by Google on the pages checked (2026-11-15) | Google Cloud model page |
| gemini-4-carbon | GA; retirement date not stated by Google on the pages checked (2026-11-15) | Google Cloud model page |
| gemini-4-argon | GA; retirement 13 October 2027 or later on Google Cloud; may remain accessible through the Gemini API after that date | Google Cloud model page |
| gemini-3.1-pro-preview | Public preview; no deprecation entry in the Gemini API changelog and no retirement date on the Google Cloud page (2026-11-15) | Gemini API changelog; Google Cloud model page |
| gemini-3-pro-preview | Shut down 9 March 2026; now points to gemini-3.1-pro-preview | Gemini API changelog |
Preview models can change or be withdrawn with less notice than GA models, which is a planning point in its own right. Track dates on the LLM Deprecation and Retirement Tracker and see the AI Model Retirement Guide.
Data handling and security
- Gemini API paid tiers: the pricing page states that Standard, Batch, Flex and Priority content is not used to improve Google's products; Contributor tier content is.
- Google Cloud controls (Gemini 4): the Argon and Carbon pages list data residency, customer-managed encryption keys (CMEK), VPC Service Controls and Access Transparency for online prediction, batch inference and context caching.
- Google Cloud (Gemini 3.1 Pro): the page states customers may use the preview in production and may process personal data under the Cloud Data Processing Addendum, subject to their Google Cloud agreement.
- Caching: implicit caching on Gemini 4 is on by default for shared prompt prefixes of 4,096 tokens or more, which matters when prompts contain sensitive material.
- Terms: on Google Cloud, Carbon is a Google Frontier Model under Section 34 of the Service Specific Terms, and the cyber variants are Google Frontier Cyber Models under Section 31.
API keys and service accounts that call these models are non-human identities. See the Google API Keys Security Guide and Zero Data Retention for AI.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| Launch stage | Argon, Carbon and Hydrogen are GA; Gemini 3.1 Pro is a public preview |
| Lifecycle runway | Argon has a stated Google Cloud retirement date of 13 October 2027 or later; no date is stated for the others |
| Sampling control | Only Gemini 3.1 Pro accepts custom temperature, top_p and top_k |
| Explicit context caching | Only Gemini 3.1 Pro |
| Output above 65,536 tokens | Gemini 4 models with long decode on v1beta1 |
| Computer use | Gemini 4 standard models |
| Region choice on Google Cloud | 3.1 Pro global only; Gemini 4 adds us and eu multi-regions, and Argon two single regions |
| Price stability | Argon and Carbon prices double on 1 December 2026; 3.1 Pro and Hydrogen list no scheduled change |
Run your own evaluation on representative prompts before switching; Google reports different benchmark suites per model.
Practitioner checklist
- Inventory every application, agent and script calling gemini-3.1-pro-preview or gemini-3-pro-preview and record the owner.
- Remove thinking_budget, MINIMAL, penalty, candidate_count and prefilled model turns before pointing code at a Gemini 4 ID, and test for HTTP 400 responses.
- Replace explicit cached_content calls with prompt ordering that puts static content of 4,096 tokens or more first.
- Re-forecast spend for 1 December 2026, when Argon and Carbon prices double.
- Set max_output_tokens deliberately: Gemini 4 counts input, thinking and output against one 1,048,576-token budget.
- Check that each project calls the exact model ID it is approved for, especially where cyber permissive variants are allowlisted.
- Store Gemini API keys and service account credentials in a secrets manager and set budget alerts.
Standards and references
- Google Cloud: Gemini 4 Argon developer guide (accessed 2026-11-15)
- Google Cloud: Gemini 4 Argon model page (accessed 2026-11-15)
- Google Cloud: Gemini 4 Carbon model page (accessed 2026-11-15)
- Google Cloud: Gemini 4 Hydrogen model page (accessed 2026-11-15)
- Google Cloud: Gemini 3.1 Pro model page (accessed 2026-11-15)
- Google AI for Developers: Gemini 3.1 Pro Preview (accessed 2026-11-15)
- Google AI Studio: Gemini 4 Hydrogen (accessed 2026-11-15)
- Google AI Studio: Gemini API pricing (accessed 2026-11-15)
- Google AI for Developers: Gemini API changelog (accessed 2026-11-15)
- Google: Gemini 4 Carbon is now available (accessed 2026-11-15)
- Google DeepMind: Gemini models (accessed 2026-11-15)
Related NHI Mgmt Group resources: Google Gemini Models, Gemini 4 Carbon, Gemini 4 Hydrogen, Frontier AI Models Compared, LLM Pricing Tracker.