# Gemini 4 Neon API: Pricing, Thinking Levels and a Sensible First Evaluation

Calculate Gemini 4 Neon token costs, separate API access from Antigravity, and evaluate a migration without changing every variable at once.

By Paras Madan | 2026-11-19


Google made Gemini 4 Neon generally available in its paid API on November 18. A model can look inexpensive per token and still make an expensive agent. The useful question with Gemini 4 Neon is not just what one response costs. It is whether the model finishes your actual task with fewer retries and less review.

I would start with a small comparison against the model already doing the job, rather than move the entire application because a new name appeared in the selector.

## What is available in the API

Google's November 18 [Gemini API release notes](https://aistudio.google.com/docs/changelog) list `gemini-4-neon` as generally available on the paid tier. They describe a unified one-million-token context window and long decoding up to one million tokens.

The supported thinking levels are `low`, `medium`, `high` and `xhigh`, with `high` as the default. That makes the thinking setting worth recording alongside the model ID in every comparison. Otherwise, a change in reasoning effort can get mistaken for a change in model quality.

Those capacity numbers are ceilings, not a reason to send every document you own. For a first evaluation, keep the input narrow enough that a person can check whether the answer actually used it correctly.

## Calculate the whole response cost

The official [Neon pricing table](https://aistudio.google.com/docs/pricing#gemini-4-neon) lists Standard rates of $3 per million input tokens and $15 per million output tokens, including thinking. There is no free tier. Cached input costs $0.15 per million tokens, with implicit caching only; explicit caching is not supported.

For a hypothetical uncached request with 100,000 input tokens and 10,000 billed output tokens, the token cost is $0.30 plus $0.15, or $0.45. If the visible answer contains 10,000 tokens but thinking adds more, that example no longer describes your bill.

Batch and Flex list $1.50 input and $7.50 output per million tokens. Compare the service conditions before using either for a time-sensitive workflow. Tool charges are separate from this simple calculation.

My suggestion is to budget an evaluation by completed tasks, not individual requests. Keep a separate allowance for retries so an unreliable workflow cannot hide behind a cheap first response.

## Do not confuse Antigravity access with API access

The [Antigravity model guide](https://antigravity.google/docs/models) lists Neon for eligible Google AI Pro plans, Ultra and Enterprise, but not Free or Plus. Its documented context window there is up to 512K tokens, distinct from the API figure.

In Antigravity, the fourth thinking option is called Extra. The CLI represents that effort as `xhigh`. Orchestrator (Neon) is a separate choice that plans with Neon and delegates implementation to Hydrogen.

Record which product you are testing. A result from an editor agent with its own tools and orchestration is not a clean measurement of a direct API call.

## Change one thing in the first trial

I would choose ten representative tasks: ordinary work, a few known failures and at least one case where the correct response is to ask for missing information. Save the original inputs and decide what a passing result looks like before running either model.

Start by preserving the prompt, tool definitions and acceptance checks. Compare the current model against Neon at a named thinking level. Then run a second comparison with a different effort setting if the first result justifies it.

For a code task, a passing answer should survive the project's tests and human review. For a document task, check important claims against the supplied evidence. A longer explanation is not an acceptance criterion.

## Make the migration decision reversible

Before expanding traffic, check structured outputs, tool arguments, error handling and actual token usage in your own application. Do not treat a successful chat response as proof that the surrounding integration is compatible.

Keep the existing route available until the comparison shows a useful gain. My threshold would be better completed work at an acceptable total cost, without a new category of failures that the team cannot detect.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn model launches and API tradeoffs into clear developer guides and technical stories.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
