Skip to content
Back to articles
Gemini 4 Neon: Pricier Than Carbon Until December 1, Then Cheaper

Gemini 4 Neon: Pricier Than Carbon Until December 1, Then Cheaper

November 20, 2026
Updated November 20, 2026

Gemini 4 Neon: Pricier Than Carbon Until December 1, Then Cheaper

Google shipped Gemini 4 Neon on November 18, and the price sheet hides a trap. At $3 per million input tokens and $15 per million output, Neon costs more per token than Gemini 4 Carbon today.

On December 1, a Gemini 4 Carbon price increase doubles it to $4/$20 and the order reverses. So if you're picking a model this week, the answer depends on the calendar.

Neon is the new middle child. Google slots it between Carbon, built for the hardest long-running tasks, and Hydrogen, the fast everyday model. It's live in the paid Gemini API under the model ID gemini-4-neon, in AI Studio, Antigravity, Gemini for business, and the Gemini app for AI Pro and Ultra subscribers.

GitHub Copilot added it the same day.

Gemini 4 Neon pricing: Gemini 4 Neon vs Carbon and Hydrogen

Here's the Standard tier per million tokens, from Google's AI Studio pricing page:

ModelInputCached inputOutput
Gemini 4 Hydrogen$1.50$0.075$7.50
Gemini 4 Neon$3.00$0.15$15.00
Gemini 4 Carbon (through Nov 30)$2.00$0.10$10.00
Gemini 4 Carbon (from Dec 1)$4.00$0.20$20.00

Run one realistic agent turn: 200K tokens in, 20K out, no cache hits. Hydrogen costs about $0.45. Neon costs $0.90; Carbon costs $0.60 today and $1.20 from December 1.

That's the whole story in four numbers. This month, Neon is 50% pricier per token than Carbon. Next month, it's 25% cheaper.

Here's the detail most launch coverage skips. Every Neon line item, including input, cached input and output, is exactly 1.5x Carbon's current rate and 0.75x Carbon's December rate. So your cache-hit ratio doesn't change the comparison at all.

What changes it is how many tokens each model burns to finish the same job.

That gives you two simple break-even rules:

  • Until November 30: Neon only wins on cost if it finishes a task using two-thirds or fewer of the tokens Carbon needs.
  • From December 1: Neon wins unless it burns more than a third more tokens than Carbon.

Google's headline claim is that Neon is about 32% cheaper per task than Carbon at high thinking. Per-task cost folds in how many tokens each model burns, so it isn't the same as the per-token rate.

Google doesn't say which Carbon price it used for that math. Treat it as directional until you've measured your own workloads.

There's no free tier for Neon. Batch and Flex run at half price ($1.50/$0.075/$7.50), and Priority costs $5.40/$0.27/$27. Argon also doubles on December 1.

Key Features and Technical Breakdown

Context and output. Per Google's Cloud docs, the window is 1,048,576 tokens, shared between input and output. A normal response caps at 65,536 output tokens.

For longer output, Google offers background long decode, which can run up to the full window minus your input. It's exposed through the Interactions API and the v1beta1 GenerateContent endpoint.

Thinking levels. You get four: LOW, MEDIUM, HIGH (default) and XHIGH. Passing MINIMAL returns an HTTP 400, and so does setting a thinking_budget. If your existing Gemini code sets either, it'll fail on Neon.

Sampling is locked. Temperature is fixed at 1.0, topP at 0.95 and topK at 64. Custom values are silently ignored. If you relied on temperature 0 for deterministic extraction, you'll need another approach, like structured output and validation.

Caching is implicit only. There's no explicit cache API. The minimum cacheable prefix is 4,096 tokens, and cache writes bill at 1.25x the input rate.

Modalities. Text, image, audio and video in. Text out, only.

Supported extras. OpenAI-compatible Chat Completions, computer use, and grounding with Google Search and Maps. Live API, tuning and RAG Engine aren't supported. Regions are global, US and EU, and the knowledge cutoff is March 2026.

Benchmarks, per Google. Google's figures put Neon at 61 on the Artificial Analysis Intelligence Index (#3) and #1 on its Coding Agent Index at 58.6. On HLE-Diamond, it's 71.3%.

The model card is more candid about coding. Neon scores 65.2% on Terminal-Bench 4.1 against Claude Opus 5.5's 65.8%, and 46.2% on FrontierCode 2.0 against Opus's 47.3%. Its clear win is long context: 95.2% on GraphWalks versus Opus 5.5's 66.8%.

Real-World Use Cases

Agent orchestration in Antigravity. Google says Neon can direct Hydrogen subagents. That's a sensible split: the pricier model plans, the cheap one executes. It's limited to paid AI Pro and Ultra subscribers.

Long-document reasoning. That GraphWalks gap is the most interesting number on the card. If you're stuffing whole repos, contracts or log dumps into context, Neon is worth a benchmark run against whatever you use now.

Copilot code changes. In Copilot, Neon is available on Pro+, Max, Business and Enterprise plans, billed at list price. GitHub says it makes fewer unrequested code changes than Carbon. Anyone who's watched an agent helpfully refactor a file they didn't ask about will get why that matters.

Drop-in for OpenAI-style clients. Because it speaks the Chat Completions format, you can point an existing OpenAI SDK client at it and A/B it without rewriting your stack.

Hard Limitations and Flaws

No safety report yet. Google lists the full safety report as coming soon. Google says Neon fell below the material-capability threshold, so it carried forward Argon's frontier-safety results. Those say biological and cyber risk cannot be ruled out. Enterprise reviewers will want the real document.

Determinism is gone. Fixed temperature hurts evals, regression tests and anything that needs repeatable output.

No rate limits published. We couldn't find them in Google's docs. Load-test before you route production traffic.

It's not the coding king. Google's own card has Opus 5.5 slightly ahead on both coding benchmarks it lists. The gap is small, but #1 coding agent depends on which index you read.

Text out only. Need images or audio back? Look elsewhere.

Verdict: Your Gemini 4 Neon Action Plan

Neon looks like the right default for long-context agent work in Google's lineup. But the timing matters more than the benchmarks.

  1. Before December 1: If you're on Carbon, your bill doubles in 11 days. Run your eval set on Neon now and compare cost per completed task, not per token.
  2. Audit your config: Strip thinking_budget, MINIMAL thinking and custom temperature settings before switching to the gemini-4-neon API, or you'll hit 400s and quiet behavior changes.
  3. Keep Hydrogen in the loop: At half Neon's rate, it's still the better pick for high-volume, simple calls.
  4. Check Copilot billing: On Copilot, Neon is billed at list price, so heavy users should watch spend.
  5. Wait on regulated workloads: Hold off until Google publishes the safety report.

The cheaper of Neon and Carbon flips on December 1. Plan the switch around that date, not the launch post.

Tools in this post

Written by

Co-founder of Tool Directory AI. Tests and reviews the AI tools listed here.

LinkedInAll posts by Sai