Key takeaways

  • Google announced Gemini 4 Neon on November 18, 2026, with rollout through the paid Gemini API, AI Studio, Gemini for business and subscriber products.
  • Launch pricing is $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. Google specifies a 1 million-token context window.
  • The official AI Studio link selects gemini-4-neon. That selector is source evidence, not an authenticated API smoke test or a guarantee of access for every account.
  • Google positions Neon between Carbon’s hardest long-running work and Hydrogen’s everyday execution. Reported rankings and cost-per-task savings are not AccessAllGPT measurements.
01

What launched, and where it is available

Google announced Gemini 4 Neon on November 18, 2026, adding a model for complex coding and knowledge work to the paid Gemini API. The launch specifies a 1 million-token context window and prices of $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. It is a managed-model rollout, not an announcement of downloadable weights.

The announcement, by Google DeepMind SVP and Google Chief AI Architect Koray Kavukcuoglu, says rollout begins in Google AI Studio, Gemini for business through its API and model dropdown, and Google Antigravity. Google AI Pro and Ultra subscribers also get access in Antigravity and the Gemini app; the closing availability section additionally names Android Studio and subscriber access in AI Studio. These are announced access paths, not accounts AccessAllGPT has tested.

Google’s linked AI Studio URL selects gemini-4-neon. We verified that identifier in the launch page’s link, not in a successful inference response. Before changing an application, confirm the callable model identifier, billing eligibility and actual availability in your own account. The launch article does not establish geographic eligibility, numeric rate limits, maximum output tokens or every supported API feature.

02

What changes for builders

Google places Neon between Carbon, which it recommends for the hardest long-horizon projects, and Hydrogen, which it describes as handling fast everyday execution. Its proposed role is complex work that happens often enough for speed and cost to matter: repository changes, analytical tasks and enterprise workflows. That is a useful candidate role, not proof that Neon should replace either neighboring model.

Google says Neon can work across spreadsheets, PDFs, charts, audio and video in its million-token context. It also describes long-running coding sessions, context compaction and an Antigravity setup in which Neon directs parallel Hydrogen subagents. Those product and harness capabilities should not be assumed to be identical to a bare API call. The launch does not provide a complete endpoint-by-endpoint modality and tool-support matrix.

The announcement describes four thinking levels, naming high as the default and xhigh as the deeper-reasoning option. We have not verified the complete parameter schema or the names of the other levels. Confirm current API documentation before copying a thinking-level value into production. Increasing reasoning effort can change both runtime and token consumption, so compare settings rather than treating the model name as the entire configuration.

03

Pricing is clear; savings per finished task are not guaranteed

The quoted input, cached-input and output prices are Google’s launch rates per million tokens. Cached input is a separate rate, not a discount on every prompt. Check cache eligibility and any separate charges before budgeting. The launch announcement does not establish a free API tier, batch discount or a universal Pro or Ultra subscription entitlement to unmetered API inference.

Google reports a 32% lower cost per task than Carbon at Neon’s default high thinking level on its discussion of Artificial Analysis results. It also cites larger savings against competing models on other benchmarks. Those are vendor-reported comparisons tied to particular evaluations and configurations, not savings measured on readers’ applications. Token rates alone cannot establish the price of a completed job when retries, tool calls and reasoning usage differ.

For a buying decision, replay representative tasks against your existing model with the same tool permissions and acceptance criteria. Record total billed usage, elapsed time, retries and whether the work actually passes. Keep the current route until Neon meets your quality threshold at an acceptable completed-task cost. AccessAllGPT has not run this comparison.

04

Read the coding scores with the harness attached

Google reports an Artificial Analysis Coding Agent Index score of 58.6, 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at the default high setting. These are figures in Google’s launch announcement; we did not independently verify current leaderboard positions or reproduce the evaluations. They support investigating the model, not promising the same success rate on a private repository.

The announcement itself illustrates why configuration matters: it reports 12.8% on SWE-sweep using the mini-SWE harness and 39.0% in Google Antigravity. That is not a controlled result we reproduced, and it should not be presented as the performance of an unspecified API call. The harness, tools and execution policy are part of the system being evaluated.

Start with a small, non-sensitive repository task, require the existing test suite to pass, and inspect the diff for unrelated edits. For document-heavy work, check source fidelity and missing qualifications rather than judging a polished summary alone. A larger context window and stronger vendor-reported scores do not remove the need for review or authorize broader agent permissions.

05

Before routing a workload to Neon

A short evaluation plan, not a benchmark we performed.

Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].

Confirm paid API eligibility and the callable identifier in your account. The announcement’s AI Studio selector is gemini-4-neon.

Record the thinking level, harness, tool permissions and token limits. Do not compare differently configured systems as if only the model changed.

Use tests and diff review for code; source fidelity and completeness checks for document work.

Compare total billed cost and elapsed time per accepted result, including retries, rather than extrapolating benchmark savings.

Primary sources

  1. Introducing Gemini 4 NeonGoogle · Reviewed: November 18 date and author; opening availability and pricing; enterprise input formats; production-ready code; efficiency and thinking levels; Gemini app analysis; closing rollout list; AI Studio link target · Retrieved · Supports: Announcement date, named access paths, 1 million-token context, $3/$0.15/$15 per-million token rates, gemini-4-neon AI Studio selector, Google-reported coding scores and harness distinction, high/xhigh discussion and vendor product positioning. Does not independently establish benchmark performance, account eligibility or successful API inference.

Limitations

No hands-on model call, authenticated access check, independent benchmark, cost measurement or live leaderboard verification. Developer-documentation retrieval required approval and was not completed; numeric rate limits, maximum output size, complete thinking-level schema and endpoint-specific feature support remain unverified. Rollout statements and performance figures are Google’s claims as retrieved on November 20. This is not biological-application coverage.

Disclosures

AccessAllGPT received no payment, private model access or vendor review for this briefing. Google did not review or endorse it. AccessAllGPT Research is operated by NeuralArc; publication-wide relationships are listed on the disclosures page.

Further AccessAllGPT guidance

  1. Model selection without leaderboard chasing
  2. Google’s Gemini 4 Carbon launch
  3. What AI models cost
  4. How this publication is made