Gemini 4 models
Gemini 4 has four models: Carbon, Neon, Hydrogen and Argon. Context window, effort levels and API prices side by side, with which to pick, as of 20 November 2026.
Pick Gemini 4 Neon for most agent and coding work, Hydrogen when speed and price matter more, and Carbon only for the hardest tasks. Google announced Neon on 18 November 2026 as the balance between Carbon, for the hardest long-horizon projects, and Hydrogen, for fast everyday execution. Argon, the fourth model, is what Google calls its efficient frontier reasoning model, with lower token spend per turn than Carbon. All four names are chemical elements.
Prices and settings below were read from Google’s docs on 20 November 2026 and change often.
The four models side by side
| Carbon | Neon | Hydrogen | Argon | |
|---|---|---|---|---|
| Model code | gemini-4-carbon |
gemini-4-neon |
gemini-4-hydrogen |
gemini-4-argon |
| Released | 10 Nov 2026 | 18 Nov 2026 | 12 Nov 2026 | 13 Oct 2026 |
| Positioning | Strongest reasoning | Frontier intelligence, fast task completion | Everyday coding and agents, speed | Efficient reasoning, fewer tokens per turn than Carbon |
| Context window | 1M tokens | 1M tokens | 1M tokens | 1M tokens |
| Effort levels | low, medium, high (default medium) | low, medium, high, xhigh (default high) | low, medium, high, xhigh (default high) | low, medium, high (default medium) |
| Input per 1M tokens | $2.00, then $4.00 from 1 Dec 2026 | $3.00 | $1.50 | $2.00, then $4.00 from 1 Dec 2026 |
| Output per 1M tokens | $10.00, then $20.00 from 1 Dec 2026 | $15.00 | $7.50 | $10.00, then $20.00 from 1 Dec 2026 |
| Free tier | No | No | Yes | No |
Prices are the Standard paid tier, in US dollars, as of 20 November 2026. All four models take text, image, audio, video and PDF input, and Google lists a maximum output of “64K / 1M tokens” for each.
What the table does not show
- Cost per task, not per token. Google says Neon at high thinking costs 32% less per task than Carbon. That is Google’s own figure, and the per-token prices above are higher for Neon than for Carbon until Carbon’s rise on 1 December 2026. Measure cost per finished task on your traces. See LLM cost optimization.
- Thinking tokens are billed. Google’s pricing page says every token processed is billed, including thinking tokens, so a higher effort level raises the bill.
- Discounts. Batch and Flex are half the Standard price (Neon $1.50 input and $7.50 output, Hydrogen $0.75 and $3.75). Priority costs more (Neon $5.40 and $27.00, Hydrogen $2.70 and $13.50). Caching is implicit only.
- Argon is not retired. The deprecations page shows “No shutdown date announced” for all four models. Released on 13 October 2026, it is the oldest of the four.
How to choose
| If you need | Start with | Why |
|---|---|---|
| A default for coding and agent loops | Neon | Google describes frontier-level intelligence with fast end-to-end task completion, and it has the xhigh level |
| High volume at the lowest price | Hydrogen | Lowest list price, a free tier, and the same four effort levels as Neon |
| The hardest reasoning or security work | Carbon | Google calls it the strongest reasoning model for complex software engineering and cybersecurity defense |
| Lower token spend than Carbon | Argon | Google says it answers faster with lower token spend per turn |
Only Neon and Hydrogen support xhigh, which Google advises reserving for hard, multi-step coding and agentic problems where high falls short. Google also shows Neon acting as an orchestrator that directs parallel Hydrogen subagents in Antigravity, a split you can copy with model routing: the larger model plans, the cheaper one executes.
Check the claims yourself
Google’s announcement quotes Artificial Analysis, Vals and Arena results for Neon. Read them at the source, with their dates and settings, before you rely on them: see how the Artificial Analysis Intelligence Index is built and what LMArena measures. A leaderboard rank says little about your prompts, so run the same task set through two candidate models and compare accuracy, cost and latency from traces. For OpenAI’s counterpart see GPT-6.2. A related Google model for closed questions is Gemini Decision.
FAQ
Which Gemini 4 model should I use?
Start with Neon for most agent and coding work and drop to Hydrogen when cost and speed matter more. Google positions Carbon as its strongest reasoning model and Hydrogen as the everyday option for coding and agentic workflows. Test on your own tasks before you commit, with an evaluation set you control.
How much does Gemini 4 cost?
As of 20 November 2026, Google's pricing page lists Neon at $3.00 input and $15.00 output per 1M tokens and Hydrogen at $1.50 and $7.50. Carbon and Argon are $2.00 and $10.00 through 30 November 2026, then $4.00 and $20.00 from 1 December 2026.
What is the context window of Gemini 4?
All four models have a 1M-token context window and accept text, image, audio, video and PDF input. See context window for what that limit means in practice.
Has Google retired Gemini 4 Argon?
Not according to Google's deprecations page, which on 20 November 2026 shows "No shutdown date announced" for gemini-4-argon, as it does for the other three models.