On 1 December 2026 the two heavier members of Google's Gemini 4 family double in price. Gemini 4 Neon, which reached general availability on 18 November, arrives at $3 per million input tokens and $15 per million output, so for many teams the more interesting question is no longer "Carbon or Argon?" but whether the middle model is now the sensible default.
Google places it squarely between Gemini 4 Hydrogen, the fast everyday worker, and Gemini 4 Carbon, the model reserved for the longest and hardest projects. The pitch is depth without Carbon's bill: complex coding, analytical reasoning and production enterprise workflows.
At a glance
| Spec |
Value |
| API ID |
gemini-4-neon |
| Released |
18 Nov 2026 (GA) |
| Input |
Text, image, audio, video, PDF and Office/OpenDocument files |
| Output |
Text |
| Token limit |
1,048,576, shared between input and output |
| Output per segment |
65,536 tokens; long decode reaches ~1M minus input |
| Thinking levels |
low, medium, high (default), xhigh |
| Knowledge cutoff |
March 2026 |
| Standard price |
$3.00 input, $0.15 cached, $15.00 output per 1M |
| Batch / Flex |
$1.50 input, $7.50 output |
| Priority |
$5.40 input, $27.00 output |
Thinking tokens are billed as output. There is no free tier, and grounding with Google Search or Maps is charged separately once the shared 5,000 monthly free requests run out ($14 per 1,000 after that). Treat all of these figures as indicative and confirm them in your own billing console.
The arithmetic after December
Through 30 November, Gemini 4 Argon and Carbon list at $2/$10. From 1 December both move to $4/$20. That flips the ranking: the newer, mid-tier model becomes the cheaper of the three upper-range options per token, while still costing twice what Hydrogen does.
Some coverage has described Neon's $3/$15 as an introductory rate that will also climb to $4/$20. Google's pricing table shows no end date for Neon and attaches the December increase only to Argon and Carbon. If your budget depends on it, check the pricing page on the day you commit rather than relying on any write-up, ours included.
Limits that change your code
Neon is less forgiving of old request shapes than Gemini 3-era models. Plan for these before switching a model string:
- One shared budget. Input and output come out of the same 1,048,576 tokens, and reasoning at high or xhigh counts toward
max_output_tokens.
- Long output is a protocol, not a flag. A single foreground response stops at 65,536 tokens. Going further requires the
v1beta1 GenerateContent endpoint or the Interactions API with continuation, plus recent SDKs (for example google-genai 2.30.1+ for GenerateContent in Python).
- Sampling is fixed. Temperature, top_p and top_k are ignored;
frequency_penalty, presence_penalty, candidate_count above 1, thinking_budget and the MINIMAL level return HTTP 400.
- Caching is implicit only. Shared prefixes of 4,096 tokens or more are cached automatically. Explicit
cached_content is not supported.
- No prefilled model turns. A conversation cannot end on a model message.
Coming from Argon specifically? Default thinking jumps from medium to high, the extra xhigh tier appears, and non-streaming output doubles from 32,768 to 65,536 tokens. Argon deployments on Google Cloud in europe-west2 or asia-southeast1 must move to the global, us or eu location.
Where you can reach it
The paid Gemini API and Google AI Studio carry it from launch, as does Gemini Enterprise. Google AI Pro and Ultra subscribers can pick it in the Gemini app, where Tasks can also run on it, and in Google Antigravity. On Google Cloud's agent platform it is GA but limited to verified enterprise customers in select regions, with Batch, Provisioned Throughput and an OpenAI-compatible Chat Completions surface; the Live API is not supported.
Press reports say it takes Argon's slot in the Gemini app's model picker. Google's own launch material says only that Neon is selectable there and is silent on Argon, so look at your own picker before telling users anything.
Is it the right tier for you?
Google's headline numbers are its own: an Artificial Analysis Intelligence Index score of 61 at xhigh, 65.2% on Terminal-Bench 4.1, and a high setting it claims is 32% cheaper per task than Carbon. Those are vendor results, not ours.
A better test takes an afternoon. Take twenty real tasks that currently go to Carbon or Argon, run them through Neon at high and again at medium, and record three things: pass rate, output tokens per task, and how often a human has to step in. If medium holds quality, you have found most of the savings. If high still falls short on your longest multi-file changes, keep Carbon for those and let Neon plan while Hydrogen subagents execute, a pattern Antigravity supports directly by passing previous_interaction_id.
More of Google's lineup lives on the Google provider page, and the comparison tool lines up Neon against rivals from other labs.