Gemini 4 Argon, Gemini 4 Carbon and Gemini 4 Hydrogen are the three Gemini 4 models Google has released, reaching general availability between 13 October and 12 November 2026. This page puts the three side by side: model IDs, release dates, access rules, token limits, thinking levels, prices with the scheduled 1 December 2026 change, tools, regions, published benchmark figures and lifecycle status. It also covers the two cyber permissive variants that Google offers to allowlisted defenders. It is written for developers choosing a model ID, buyers estimating cost and security teams planning access and data controls. Models are listed alphabetically (Argon, Carbon, Hydrogen); order is never a ranking. Every fact comes from Google's own documentation, model card and announcements, cited in the sources list.
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
- Argon and Carbon share a price; Hydrogen lists lower rates. Argon and Carbon list $2 input and $10 output per million tokens until 30 November 2026, then $4 and $20; Hydrogen lists $1.50 and $7.50 with no scheduled change.
- Carbon is built on Argon. The Gemini 4 Carbon model card states Carbon is based on Gemini 4 Argon and refers to the Argon card for architecture and training details.
- All three share a unified 1,048,576-token limit. Input and output share one window, a foreground segment produces up to 65,536 output tokens, and long decode extends output up to the full limit.
- Default thinking differs. Argon and Carbon default to MEDIUM on Vertex AI; Hydrogen defaults to HIGH and adds an XHIGH level.
- Access differs. Argon is publicly available on Vertex AI, Carbon on Vertex AI is for verified enterprise customers in select regions, and Hydrogen is the only one of the three with a free tier on the Gemini API.
What are Gemini 4 Argon, Carbon and Hydrogen?
Gemini 4 Argon was announced on 30 September 2026 and reached general availability on 13 October 2026. The Gemini API changelog entry for that date calls it "our most capable model built for complex coding, agentic, math, and science tasks", and the Google AI Studio models page describes it as "Our flagship reasoning model with a shared 1M context window and long decodes up to 1M tokens".
Gemini 4 Carbon was released on 10 November 2026. Google's announcement calls it "our strongest model", with frontier performance in cybersecurity, long-horizon agentic workflows and scientific reasoning, and states that Argon "remains the ideal choice for the majority of software engineering and knowledge work". Carbon targets harder tasks and uses more tokens to explore and verify.
Gemini 4 Hydrogen was released on Vertex AI on 12 November 2026. Its Vertex AI page describes it as "built for speed and efficiency across everyday coding and agentic workflows", and the AI Studio model page says it brings "cutting-edge intelligence to everyone".
All three are proprietary: Google publishes no weights and states no parameter count. For the full Gemini line-up, see the Google Gemini Models hub; for single-model detail, see Gemini 4 Argon, Gemini 4 Carbon and Gemini 4 Hydrogen.
Every model at a glance
As of 2026-11-15. Source: Vertex AI model versions page and Vertex AI model pages.
| Model | API ID | Released | Status |
|---|---|---|---|
| Gemini 4 Argon | gemini-4-argon | 2026-10-13 | Generally available |
| Gemini 4 Argon Cyber Permissive | gemini-4-argon-cyber-permissive | 2026-10-13 | Allowlist only (Fairwind Program); shorter availability period |
| Gemini 4 Carbon | gemini-4-carbon | 2026-11-10 | Generally available; shorter availability period |
| Gemini 4 Carbon Cyber Permissive | gemini-4-carbon-cyber-permissive | 2026-11-10 | Allowlist only (Fairwind Program); shorter availability period |
| Gemini 4 Hydrogen | gemini-4-hydrogen | 2026-11-12 | Generally available (Stable on the Gemini API); shorter availability period |
Open weights: none. Licence: proprietary. Parameters: not stated by Google. The Vertex AI model versions page states that models in the shorter availability period stay active until a replacement launches and a retirement date is announced. For how Gemini 4 relates to the previous generation, see Gemini 4 vs Gemini 3.1 Pro.
When was each model released?
As of 2026-11-15. Dates from Google's announcements, the Gemini API changelog and the Vertex AI model versions page, shown chronologically.
| Date | Event | API ID |
|---|---|---|
| 2026-09-30 | Gemini 4 Argon announced; gradual rollout to paid API customers and Google AI Ultra subscribers first; cybersecurity experts already using it through the Fairwind program | gemini-4-argon |
| 2026-10-13 | Gemini 4 Argon generally available on the Gemini API and Vertex AI; Argon Cyber Permissive released on Vertex AI | gemini-4-argon, gemini-4-argon-cyber-permissive |
| 2026-11-10 | Gemini 4 Carbon announced and released on Vertex AI, with Carbon Cyber Permissive for Fairwind Program members | gemini-4-carbon, gemini-4-carbon-cyber-permissive |
| 2026-11-12 | Gemini 4 Hydrogen released on Vertex AI | gemini-4-hydrogen |
| 2026-12-01 | Scheduled: Argon and Carbon introductory prices end and double | gemini-4-argon, gemini-4-carbon |
How much does each model cost?
As of 2026-11-15. USD per million tokens, Standard tier, from the Gemini API pricing page. Output prices include thinking tokens. No prompt-length tiers are listed for these models.
| Model | Input | Cached input | Cache write | Output | Batch (input / output) | Notes stated by provider |
|---|---|---|---|---|---|---|
| Gemini 4 Argon | $2.00 ($4.00 from 2026-12-01) | $0.10 ($0.20 from 2026-12-01) | No separate charge listed (implicit caching only) | $10.00 ($20.00 from 2026-12-01) | $1.00 / $5.00 ($2.00 / $10.00 from 2026-12-01) | No free tier; Priority $3.60 / $18.00; Contributor tier (invite-only) $1.00 / $5.00 |
| Gemini 4 Carbon | $2.00 ($4.00 from 2026-12-01) | $0.10 ($0.20 from 2026-12-01) | No separate charge listed (implicit caching only) | $10.00 ($20.00 from 2026-12-01) | $1.00 / $5.00 ($2.00 / $10.00 from 2026-12-01) | No free tier; Priority $3.60 / $18.00 |
| Gemini 4 Hydrogen | $1.50 | $0.075 | No separate charge listed (implicit caching only) | $7.50 | $0.75 / $3.75 | Free tier available; Priority $2.70 / $13.50; Contributor tier (invite-only) $0.75 / $3.75 |
Flex tier: Flex prices match Batch prices for all three models. Priority prices for Argon and Carbon also double on 1 December 2026, to $7.20 input and $36.00 output.
Grounding: on paid tiers, Grounding with Google Search and with Google Maps each include 5,000 free requests a month shared across Gemini 3 and later models, then $14 per 1,000. Grounding is not available on Hydrogen's free tier.
Content use on free and Contributor tiers: the pricing page states that content on Hydrogen's free tier, and on the Contributor tier for Argon and Hydrogen, may be used to improve Google's products. For cost guardrails, see How to set spend limits on LLM API keys and the LLM Pricing Tracker.
What changed between models?
Only differences Google documents are listed. Google has not published a changelog between the three models; the points below come from each model's documentation and announcement.
- Argon (first Gemini 4 model): the announcement states the output limit rose to 1 million tokens, up from 64,000, and that cached input is discounted by 95%. Vertex AI lists thinking levels LOW, MEDIUM (default) and HIGH, and an implicit caching minimum of 4,096 tokens.
- Argon to Carbon: Carbon is based on Argon (model card). Carbon keeps the same thinking levels and default, the same token limits and the same list price. Google's announcement reports, against Argon, a per-attempt breach rate of 0.8% (down from 3.1%), automated cyberattack query success of 2.17% (down from 9.86%) and 92% of 375 agentic honesty tasks passed (up from 70%). The model card reports a 3.50 percentage point gain in text-to-text safety against Argon. On Vertex AI, Carbon's regions are global, us and eu, against five listed regions for Argon.
- Hydrogen: adds an XHIGH thinking level and defaults to HIGH. The AI Studio model page links a "What's new in Gemini 4 Hydrogen" page covering API changes from Gemini 3.8 Flash. Hydrogen lists lower prices and a free tier.
- All three: setting thinking_budget or thinking_level MINIMAL returns HTTP 400 on Vertex AI, and custom temperature, top_p and top_k values are ignored.
For thinking and effort background, see Reasoning Models Explained.
What workloads does Google position each model for?
This section records the provider's own positioning wording; it is not an assessment.
| Model | Positioning stated by provider | Related notes stated by provider |
|---|---|---|
| Gemini 4 Argon | "Remains the ideal choice for the majority of software engineering and knowledge work" (Carbon announcement) | Vertex AI: frontier model for coding, finance, legal, cybersecurity and science |
| Gemini 4 Carbon | "Our strongest model" for harder tasks such as multi-step coding, debugging and open-ended research (announcement) | Uses more tokens to explore and verify |
| Gemini 4 Hydrogen | "Built for speed and efficiency across everyday coding and agentic workflows" (Vertex AI) | AI Studio: "cutting-edge intelligence to everyone" |
| Cyber Permissive variants | Defensive cybersecurity work such as vulnerability discovery and patch generation (Vertex AI) | Allowlist through the Fairwind Program; default thinking HIGH; computer use not offered |
Worked cost example
Using Standard list prices, 1,000 agent runs a day at 200K input and 20K output tokens each uses 200 million input and 20 million output tokens a day. On Argon or Carbon at introductory prices that is $400 input plus $200 output, $600 a day; from 1 December 2026 it becomes $1,200 a day. On Hydrogen it is $300 plus $150, $450 a day. Carbon's announcement states that it uses more tokens to explore and verify, and Hydrogen defaults to a higher thinking level than Argon, so measured output tokens can differ from this estimate. Re-run the arithmetic with your own token counts.
What is the context window of each model?
As of 2026-11-15. From the Vertex AI model pages and the AI Studio model pages.
| Model | Context window | Max output | Knowledge cutoff | Input | Output | Reasoning controls |
|---|---|---|---|---|---|---|
| Gemini 4 Argon | 1,048,576 tokens, shared by input and output | 65,536 per streaming segment (32,768 non-streaming); up to 1,048,576 with long decode | Not stated by Google on the pages checked (2026-11-15) | Text, image, audio, video, PDF | Text | Thinking LOW, MEDIUM (default), HIGH |
| Gemini 4 Carbon | 1,048,576 tokens, shared by input and output | Up to 1,048,576 with long decode | March 2026 | Text, image, audio, video; Office and OpenDocument files | Text | Thinking LOW, MEDIUM (default), HIGH |
| Gemini 4 Hydrogen | 1,048,576 tokens, shared by input and output | 65,536 by default; up to 1,048,576 with long decode | March 2026 | Text, image, audio, video, PDF | Text | Thinking LOW, MEDIUM, HIGH (default), XHIGH |
Long decode runs on the Interactions API and the v1beta1 GenerateContent API on Vertex AI. Hydrogen's Vertex AI page also lists media limits of up to 3,000 images or documents per prompt and up to 10 videos per prompt. See Context Windows Explained.
Capabilities and tools compared
As of 2026-11-15. From the Vertex AI model pages, with Hydrogen tools also confirmed on its AI Studio page.
| Model | Function calling and structured outputs | Computer use tool | Fine-tuning | Cloud IDs stated by provider |
|---|---|---|---|---|
| Gemini 4 Argon | Function calling, structured output, grounding, code execution, URL context | Supported (not on the Cyber Permissive variant) | Tuning not offered on Vertex AI | gemini-4-argon (Gemini API and Vertex AI) |
| Gemini 4 Carbon | Function calling, structured output, grounding, code execution, URL context | Supported (not on the Cyber Permissive variant) | Tuning not offered on Vertex AI | gemini-4-carbon (Gemini API and Vertex AI) |
| Gemini 4 Hydrogen | Function calling, structured output, grounding with Google Search and Maps, code execution, URL context, file search | Supported | Tuning not offered on Vertex AI | gemini-4-hydrogen (Gemini API and Vertex AI) |
The Gemini Live API is listed as unavailable for all three models. RAG Engine on Vertex AI is also listed as unavailable for all three; grounding and file search cover retrieval use cases. For agent tooling risks, see Computer-Use and Browser Agents.
Where is each model available?
| Model | Availability stated by provider |
|---|---|
| Gemini 4 Argon | Gemini API and Google AI Studio (paid); Vertex AI publicly available in global, us, eu, europe-west2 and asia-southeast1; Provisioned Throughput, batch and pay-as-you-go (standard and priority) |
| Gemini 4 Carbon | Gemini API and Google AI Studio (paid); Gemini for business; rolling out to Google AI Ultra subscribers in the Gemini app, Google Antigravity and AI Mode in Search; Vertex AI for verified enterprise customers in select regions (global, us, eu) |
| Gemini 4 Hydrogen | Gemini API and Google AI Studio (free and paid tiers; Batch, Flex and Priority); Vertex AI in global, us and eu; Provisioned Throughput, batch and pay-as-you-go |
The Cyber Permissive variants run in global and us (Carbon) or global, us and eu (Argon). Google's Carbon announcement also describes a free six-month programme giving eligible scientists Argon and Carbon access in Antigravity.
Model cards and system cards
| Model | Model card or page | System card | Technical report |
|---|---|---|---|
| Gemini 4 Argon | Vertex AI model page; model card referenced by the Carbon model card | Not stated by Google on the pages checked (2026-11-15) | Not stated by Google on the pages checked (2026-11-15) |
| Gemini 4 Carbon | Model card (November 2026) | Gemini 4 Carbon Safety and Security Report, linked from the announcement | Not stated by Google on the pages checked (2026-11-15) |
| Gemini 4 Hydrogen | AI Studio model page, which links a model card | Not stated by Google on the pages checked (2026-11-15) | Not stated by Google on the pages checked (2026-11-15) |
This page does not summarise card contents beyond the facts cited above.
Benchmarks
As of 2026-11-15. Figures are provider-reported, as stated in the text of each Google announcement or card. Suites differ between announcements, so rows are not directly comparable.
| Model | Provider-reported benchmarks | Independent leaderboard results |
|---|---|---|
| Gemini 4 Argon | DeepSWE v1.1 77.9%; AutomationBench 51.3%; LVBench 91.7%; CWE-bench v1 68% (announcement, 2026-09-30) | Not checked for this page |
| Gemini 4 Carbon | Project Fortify (Google internal) 100% recall on scanning and 95% precision on triage; ranks stated without scores for CursorBench 4.0, WANDR and AutomationBench (announcement, 2026-11-10) | Not checked for this page |
| Gemini 4 Hydrogen | Not stated by Google on the pages checked (2026-11-15) | Not checked for this page |
Deprecations and retirements
As of 2026-11-15. Source: Vertex AI model versions page. Once a retirement is scheduled for a model in the shorter availability period, Google commits to at least 45 days to migrate.
| Model | API ID | Notice or deprecation date | Retirement or shutdown date | Stated replacement |
|---|---|---|---|---|
| Gemini 4 Argon | gemini-4-argon | Active | 2027-10-13 or later | None stated |
| Gemini 4 Argon Cyber Permissive | gemini-4-argon-cyber-permissive | Active | No retirement date announced | None stated |
| Gemini 4 Carbon | gemini-4-carbon | Active | No retirement date announced | None stated |
| Gemini 4 Carbon Cyber Permissive | gemini-4-carbon-cyber-permissive | Active | No retirement date announced | None stated |
| Gemini 4 Hydrogen | gemini-4-hydrogen | Active | No retirement date announced | None stated |
The Gemini API changelog lists no deprecations for these models as of the date checked. See the AI Model Retirement Guide and the LLM Deprecation and Retirement Tracker.
Data handling and security
On Vertex AI, each of the three model pages lists data residency, customer-managed encryption keys (CMEK), VPC Service Controls (VPC-SC) and Access Transparency (AXT) for online prediction, batch inference and context caching. On the Gemini API, the pricing page states that free-tier and Contributor-tier content may be used to improve Google's products, which matters for Hydrogen's free tier; confirm paid-tier terms in your own agreement.
The Cyber Permissive variants are released with fewer cyber safeguards to allowlisted defenders, so treat the credentials and projects that can call them as sensitive. Google's Carbon announcement also states that prompt injection robustness "remains an area for continued improvement". Gemini API keys and Vertex AI service accounts are the controls that matter most: see the Google Gemini API and AI Studio Security Guide, the Google Vertex AI Security Guide and the Google API Keys Security Guide.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| List price per token | $2/$10 for Argon and Carbon until 2026-11-30, then $4/$20; $1.50/$7.50 for Hydrogen |
| Free tier | Hydrogen only |
| Vertex AI access | Argon and Hydrogen publicly available; Carbon for verified enterprise customers in select regions |
| Region choice on Vertex AI | Argon lists europe-west2 and asia-southeast1 in addition to global, us and eu |
| Default thinking level | MEDIUM on Argon and Carbon; HIGH on Hydrogen, which also offers XHIGH |
| Knowledge cutoff stated | March 2026 for Carbon and Hydrogen; not stated for Argon |
| Lifecycle runway | Argon has a stated retirement of 2027-10-13 or later; Carbon and Hydrogen have no date announced |
Run your own evaluation on representative prompts before choosing; Google's published figures use different suites per announcement.
Practitioner checklist
- Pin production code to an explicit model ID (gemini-4-argon, gemini-4-carbon or gemini-4-hydrogen) and record it in the AI inventory.
- Budget for the 1 December 2026 price change on Argon and Carbon, including Batch, Flex and Priority tiers.
- Keep production and sensitive data off Hydrogen's free tier and the Contributor tier.
- Remove thinking_budget and MINIMAL thinking settings from code before moving to Gemini 4, and stop relying on custom sampling values.
- Restrict which projects and service accounts can call the Cyber Permissive variants, and log their use.
- Turn on CMEK, VPC-SC and data residency settings on Vertex AI where your policy requires them.
- Measure output and thinking tokens per task on each model rather than relying on list prices.
- Watch the Vertex AI model versions page for retirement dates; Google commits to at least 45 days' notice.
Standards and references
- Google Cloud: Model versions and lifecycle (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: Gemini API pricing (accessed 2026-11-15)
- Google: Gemini API models (accessed 2026-11-15)
- Google: Gemini 4 Hydrogen model page, Gemini API (accessed 2026-11-15)
- Google: Gemini API changelog (accessed 2026-11-15)
- Google DeepMind: Gemini 4 Carbon model card (accessed 2026-11-15)
- Google: Gemini 4 Carbon announcement (accessed 2026-11-15)
- Google: Gemini 4 Argon announcement, Spanish edition (accessed 2026-11-15)
Related NHI Mgmt Group resources: Google Gemini Models, Gemini 4 vs Gemini 3.1 Pro, Gemini Flash and Flash-Lite Versions Compared, Frontier AI Models Compared, LLM Pricing Tracker.