Gemini 4 Neon steps up, with a higher token cost

Google makes Gemini 4 Neon available to the general public via its API and Studio offerings. The model targets complex tasks, but its pricing requires close monitoring of token consumption.

Google made Gemini 4 Neon generally available on November 18, 2026, through its Gemini API, with a clear positioning: a paid model for coding, research, and science tasks, priced at 3 dollars per million input tokens and 15 dollars per million output tokens. The bet is simple to state, less simple to monetize: Neon promises fewer steps and fewer tokens used, but its unit price is higher than some neighboring models. For product teams and developers, the bill will therefore mainly depend on the actual reasoning volume, not the marketing around the word “intelligent.”

Un modèle taillé pour les tâches complexes, pas pour le décor

Neon, a model from the Gemini 4 family, sits between Hydrogen and Carbon according to Google's launch guide. The company describes it as suitable for high-frequency and complex tasks, with uses in software engineering, document work, and science. In practice, Google assigns it debugging capabilities across multiple repositories, multi-file refactoring, and transforming dense documents, tables, or tool results into structured deliverables.

This upgrade follows Argon, released on October 13, 2026, Carbon on November 10, 2026, and Hydrogen on November 12, 2026. Neon thus becomes the fourth Gemini 4 model in this sequence. Carbon remains Google's recommendation for particularly difficult problems in architecture, debugging, science, and security, while Hydrogen remains the default choice for routine coding and agent workflows.

Le vrai sujet : le prix facial ou le coût final ?

Neon's pricing is higher than Hydrogen's, which is charged at 1.50 dollars per million input tokens and 7.50 dollars per million output tokens, with free API access on the latter. Carbon and Argon are priced at 2 dollars per million input tokens and 10 dollars output until November 30, 2026, before rising to 4 dollars and 20 dollars respectively from December 1, 2026. In this context, Google pushes a performance-by-reasoning logic, not low price.

However, the calculation does not stop at the gross price. Neon's output includes reflection tokens, and Google also charges for certain internal tool instructions as well as function declarations used with grounding. Grounding via Google Search is additionally subject to separate pricing after a shared quota of 5,000 free requests per month. In other words, the model can cost less in use if it truly cuts the number of calls and steps, but it can also increase the bill if prompts are long or tools are called in cascade.

Un grand conteneur, mais pas un puits sans fond

Google sets the shared window between input, reasoning, and output at 1,048,576 tokens. Synchronous responses are capped at 65,536 tokens by default, although background execution or client-side continuation can extend generation within the overall budget. The field that sets the output thus also caps the thinking tokens: with too low a limit, the response may be cut off before completion. Not ideal when you want an “advanced” model and you throttle it like a scooter on the sidewalk.

Google offers four reasoning levels: low, medium, high, and xhigh, with high enabled by default. The company recommends testing all four levels and choosing the lowest that meets the expected quality. The advice is pragmatic: xhigh is only justified for difficult multi-step problems when high is no longer sufficient. For teams deploying automated support, document analysis, or code assistants, configuration matters almost as much as the chosen model.

Ce que Neon accepte, et ce qu’il ne fait pas

Neon supports text, images, video, audio, and PDFs as input, but generates text only. The model also accepts code execution, computer usage, function calls, file search, structured outputs, URL context, and grounding via Google Search and Google Maps. However, it does not offer image generation, audio generation, or Live API.

Google states that Antigravity, its managed agent, can operate with Neon alone or combined with Hydrogen in multi-model workflows. The provider does not require a query schema change to switch from Carbon, Argon, or Hydrogen to Neon, but default values change: Neon starts at high and can go up to xhigh, while Carbon and Argon start at medium and cap at high. Migrations from Gemini 3 and earlier versions are more delicate, notably due to the removal of certain explicit cache fields and sampling parameters, as well as preserving previous turns and thought signatures in client-managed histories.

Key points

  • November 18, 2026: general availability of Gemini 4 Neon.
  • 3 dollars / million input tokens, 15 output.
  • Shared window of 1,048,576 tokens.
  • High enabled by default, xhigh available.
  • Google recommends testing all four reasoning levels.
  • Carbon remains recommended for the most difficult cases.

By the numbers

  • 5,000 requests — free monthly quota for Google Search grounding.
  • 65,536 tokens — default synchronous response cap.
  • 1,048,576 tokens — total shared input/reasoning/output window.
  • 15 dollars — cost per million output tokens for Neon.
  • November 30, 2026 — end of promotional pricing for Carbon and Argon.

To read