> ## Content Index
> Fetch the complete content index at: https://www.tao.media/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google Launches Gemini 4 Neon for Complex Coding at $3 Input and $15 Output per Million Tokens
- URL: https://www.tao.media/google-launches-gemini-4-neon-for-complex-coding-at-3-input-and-15-output-per-million-tokens/
- Published: 2026-11-19T16:13:52.000Z
- Updated: 2026-11-19T16:13:52.000Z
- Description: The new model pairs long-context reasoning with a Neon-and-Hydrogen workflow that divides planning from routine execution.
- Author: Bart Hillerich
- Tags: Google, AI, News

Google has launched [Gemini 4 Neon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/?ref=tao.media), a model for complex coding, analytical reasoning, and enterprise work priced at $3 per million input tokens and $15 per million output tokens. It sits between Carbon, which Google recommends for its hardest long-horizon projects, and Hydrogen, its model for fast everyday execution.

The November 18 rollout brings another option for developers running demanding tasks frequently enough that execution time and token spending matter alongside reasoning quality. Google reports that Neon at its default high thinking level costs 32% less per task than Carbon on the Artificial Analysis Intelligence Index comparison.

Neon is [rolling out across the Gemini API](https://x.com/GoogleDeepMind/status/2123167884411077030?ref=tao.media), Google's development tools, Gemini for business, the Gemini app, and Antigravity. In Antigravity, a new Orchestrator pairing lets Neon plan a task while Hydrogen handles routine implementation, with control returning to Neon when execution runs into difficulty.

![Neon is best in class on Terminal-Bench Science, RiemannBench, and HLE-Diamond](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/HXb_xtqWIAAVtPL.png)

## A Shared Context Budget for Inputs, Reasoning, and Output

Neon is adapted from Carbon, according to its [DeepMind model card](https://deepmind.google/models/model-cards/gemini-4-neon/?ref=tao.media). It accepts text, images, audio, and video, allowing a workflow to combine documents, spreadsheets, charts, and other materials before producing a structured response.

The model's 1,048,576-token context window is a unified budget: input, internal reasoning, and generated output all occupy the same window. A larger document collection therefore leaves less room for reasoning and the final answer, rather than sitting alongside a separate million-token output allowance.

Developers call the model as `gemini-4-neon`. Its [API documentation](https://aistudio.google.com/docs/whats-new-gemini-4-neon?ref=tao.media) lists four thinking levels: `low`, `medium`, `high`, and `xhigh`. The default is `high`. Lower levels reduce latency and token use; `xhigh` allocates more reasoning to difficult engineering, analytical, and scientific problems.

Synchronous responses support up to 65,536 tokens per hop. Longer generation uses background execution or client-side continuation, extending across multiple hops within the remaining unified context budget. The `max_output_tokens` setting caps reasoning and output together, so developers need to leave room for both rather than budgeting only for the visible answer.

For migrations from another Gemini 4 model, Google says the request schema stays the same. Teams still need to check the thinking setting: Carbon defaults to medium, while Neon defaults to high. Neon also supports the Antigravity agent in Gemini Managed Agents, either alone or paired with Hydrogen.

## Coding and Knowledge-Work Results

Google reports that Neon scores 61 at `xhigh` on the Artificial Analysis Intelligence Index, ranking third overall. That result uses a higher reasoning setting than the default-high configuration behind Google's 32% cost-per-task comparison with Carbon.

For agentic work, Google reports first-place results of 58.6 on the Artificial Analysis Coding Agent Index and 56.3% on Zapier's AutomationBench. The company also reports second place on the Vals Index at 70.9%, including first-place rankings on Legal Research Bench and Tax Agent Bench.

These evaluations cover different kinds of work. Coding-agent tests assess a model operating through a software environment, while legal, tax, and automation tasks test its ability to use information and tools toward a practical result. Google says Neon also ranks second on steerability in Agent Arena, measuring how well an agent adapts to corrections during a task.

![Gemini 4 Neon benchmark chart](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/gemini-4-neon__benchmark__secben.width-1600.format-webp.webp)

[Gemini 4 Neon benchmark chart | Google Blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/?ref=tao.media)

Separately, [Cursor reports](https://x.com/cursor%5Fai/status/2123198346802450926?ref=tao.media) that Neon at default effort scores 63.7% on CursorBench 4.1, compared with 57.3% for Gemini 4 Hydrogen, a gain of 6.4 percentage points. Cursor has made Neon available in its editor and characterizes the result as a new cost-performance frontier on its benchmark.

Google's reliability claims include its lowest fabrication rate yet in internal Antigravity testing. That is an internal evaluation, distinct from the named external indices. The company also reports improvements in avoiding shortcuts and staying within the requested scope during multi-step work.

## Neon Plans While Hydrogen Executes

The [Antigravity integration](https://antigravity.google/blog/gemini-4-neon-in-google-antigravity/?ref=tao.media) targets repository-level debugging, multi-file refactoring, secure code generation, enterprise analysis, and scientific reasoning. Google says Neon reduces redundant file reads and handles context compaction more reliably during long sessions with many tool calls. Context compaction condenses earlier work so an agent can continue without carrying every previous interaction in full.

Orchestrator divides a session between two models rather than requiring users to switch manually. Neon handles initial planning, architecture, and difficult reasoning; Hydrogen takes routine code edits, package updates, and test runs. When Hydrogen encounters friction, Neon inspects the current state, adjusts the plan, and resumes the workflow.

The arrangement conserves usage quota by reserving Neon's reasoning for the steps that need it. It also gives developers a choice between running Neon throughout a task and using it to supervise a faster execution model.

Security work spans code generation and repairing existing vulnerabilities. Google reports fewer vulnerabilities in generated code and stronger root-cause patching, while distributing Neon for cyber's dual-use offensive cyberdefense capabilities through its [Fairwind Program](https://deepmind.google/fairwind-program/?ref=tao.media). That version is scheduled to reach trusted defenders in the days following the launch.

## Pricing and Access

Neon's standard API rates are $3 per million input tokens and $15 per million output tokens. Cached input costs $0.15 per million tokens, a 95% discount on the input rate. The API uses implicit caching; the discounted rate applies to cache hits rather than all repeated material automatically qualifying.

The paid Gemini API supports Neon through both `generateContent` and the Interactions API. Developers can also access it through Google AI Studio and Android Studio, while Gemini for business offers API access and model selection.

Google AI Pro and Ultra subscribers receive access in AI Studio and the Gemini app. For paid Pro and Ultra subscribers, Antigravity's rollout covers its desktop app, CLI, and IDE extensions, including the Orchestrator pairing of Neon and Hydrogen.