← All articles

Gemini 4 Neon: API Pricing, Access and Migration Checks

Google launched Gemini 4 Neon for coding and agent work. Compare API pricing, thinking levels, Antigravity access and the migration limits to check.

Google launched Gemini 4 Neon on November 18 for complex coding, analysis and agent work. For a team considering it, the useful questions are which product exposes it, what a completed task costs and whether an existing integration needs changes. This guide separates the launch claims from practical checks you can run yourself. Google’s announcement

What changed with Gemini 4 Neon?

Google positions Neon between Carbon’s deeper long-horizon work and Hydrogen’s everyday execution. The launch also includes Neon as an Orchestrator in Antigravity. Google’s reported benchmark results are reasons to evaluate it, not proof that your own workload will improve. Model positioning

Keep a baseline from the model you already use. Choose a real task with a clear finish line: a bounded repository fix, a sourced document summary or a structured extraction that someone can verify. Use the same inputs and acceptance criteria for both models. Record unsuccessful attempts and corrections as well as the best result.

What does the API cost?

Google announced these rates per million tokens:

Token category Announced rate
Input $3
Cached input $0.15
Output $15

These are the launch rates, not a flat subscription price. Pricing announcement

As an illustrative calculation, 100,000 uncached input tokens and 20,000 output tokens at those rates total $0.60. That is a token-only estimate: inspect actual billable usage and any additional services, currency conversion or taxes in your chosen account. Repeated attempts and long reasoning can change the cost of finishing a task.

For a pilot, set a budget for completed, accepted work. A cheaper single response can still be expensive if a person must correct it repeatedly. Compare the cost and review time needed to reach the same usable result.

Where can you access Neon?

The launch lists the Gemini API paid tier, Gemini for business, paid Google AI Pro and Ultra access in Antigravity, and Pro/Ultra access in the Gemini app. A separate cyber variant is being offered to eligible Fairwind defenders. Access routes

Check access in the exact workspace and product you intend to use. Cloud’s model page notes verified enterprise customers and selected regions; do not interpret a launch listing as universal access. Verify API billing separately from consumer subscription access. Cloud model specifications

Thinking and context limits need separate checks

Cloud documents four thinking levels: LOW, MEDIUM, HIGH and XHIGH, with HIGH as the default. It also specifies a unified 1,048,576-token limit shared by input and output, with thinking and function calls counted in output. A single foreground segment produces up to 65,536 output tokens; longer output uses supported continuation paths. Neon API limits

The Antigravity team states up to 512K tokens in Antigravity. Its Orchestrator pairs Neon’s planning and review with Hydrogen’s implementation work. Paid Pro and Ultra include that mode; the team’s table says Enterprise gets Neon through Gemini for business but does not include Orchestrator. Official Antigravity team announcement

Choose effort intentionally. Start from the documented default, then compare a lower level on straightforward work and a higher level on a task that warrants it. Measure whether the extra reasoning changes the accepted result. Do not assume that a larger advertised context limit means your selected client will retain or use the same amount of information.

What should an API migration review cover?

Google’s migration documentation identifies changes beyond replacing the model name. Use the current documentation for the API and SDK you deploy:

  1. Confirm the target and client. The model ID is gemini-4-neon. Google’s getting-started guide specifies Python google-genai 2.34.0 or later and JavaScript/TypeScript @google/genai 2.32.0 or later.
  2. Review reasoning configuration. Replace older thinking-budget or minimal-effort settings with a supported thinking level.
  3. Check generation controls. Custom temperature, top-p and top-k values are ignored; revisit evaluations that assumed those settings controlled the output.
  4. Preserve tool-turn state. Review thought-signature handling, including parallel function calls, rather than stripping or duplicating signatures in your adapter.
  5. Recalculate token budgets. Account for the combined input/output limit and verify any long-output continuation behavior with your actual API version.

Official Gemini 4 migration guide

Use a test environment for that review. Include one normal tool call, one tool failure and a multi-turn request. Compare the model’s final statement with the real tool result. A passing request or a confident completion message is not enough evidence that a file, record or external action changed correctly.

A practical first evaluation prompt

Adapt this original prompt to a repository you can inspect:

Investigate [specific defect] in this test repository. Identify the relevant files and evidence, then propose a small fix and a meaningful verification step. Keep the change within that scope. Report what you actually checked, what failed and what remains uncertain. Do not commit, deploy or modify external systems.

For document work, replace the defect with a specific question and require source references for important claims. Keep sensitive production information out of the first pilot until your data-handling requirements and chosen access route have been reviewed.

Our Gemini 4 Hydrogen guide covers the earlier model launch. For help choosing and validating an AI workflow, Milan’s AI consulting service can help define the task, integration and acceptance checks. Browse recent engineering projects for related delivery work.

Frequently asked questions

What is Gemini 4 Neon for?

Google positions Neon for complex coding, analysis and agent workflows, balancing reasoning depth with execution efficiency.

What thinking levels does Neon support?

Neon supports LOW, MEDIUM, HIGH and XHIGH, with HIGH as the default. Test the effort level against your own task and cost target.

Is the context limit the same in every product?

No. Cloud documentation describes a unified 1,048,576-token input-and-output limit, while the Antigravity team states up to 512K tokens in Antigravity.

Should I replace my current model immediately?

Run a small comparison first. Check correctness, tool behavior, completed-task cost and review time before changing a production workflow.

FAQ

Frequently asked questions

What is Gemini 4 Neon for?

Google positions Neon for complex coding, analysis and agent workflows, balancing reasoning depth with execution efficiency.

What thinking levels does Neon support?

Neon supports LOW, MEDIUM, HIGH and XHIGH, with HIGH as the default. Test the effort level against your own task and cost target.

Is the context limit the same in every product?

No. Cloud documentation describes a unified 1,048,576-token input-and-output limit, while the Antigravity team states up to 512K tokens in Antigravity.

Should I replace my current model immediately?

Run a small comparison first. Check correctness, tool behavior, completed-task cost and review time before changing a production workflow.

Need help with AI model evaluation and integration?

Turn the idea into a working system.