# Gemini 4 Neon

> Source: https://aiwiki.ai/wiki/gemini_4_neon
> Updated: 2026-11-22
> Categories: AI Models, Google DeepMind, Large Language Models, Multimodal AI, Reasoning Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Gemini 4 Neon." aiwiki.ai, 22 Nov 2026. https://aiwiki.ai/wiki/gemini_4_neon
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Gemini 4 Neon is a [multimodal](https://aiwiki.ai/wiki/multimodal_model) [reasoning model](https://aiwiki.ai/wiki/reasoning_models) from [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind), announced on November 18, 2026 for software engineering, analytical work, and scientific tasks.[1] Google adapted it from [Gemini 4 Carbon](https://aiwiki.ai/wiki/gemini_4_carbon).[2] The stable API identifier is `gemini-4-neon`; the model accepts text, images, audio, video, and PDFs and produces text.[3]

Google positions Neon between Carbon for demanding extended work and Hydrogen for everyday execution. These are workload distinctions rather than a guarantee that one model wins every comparison.[4] Neon also serves as a planner in [Google Antigravity](https://aiwiki.ai/wiki/antigravity), where an Orchestrator workflow can delegate implementation to Hydrogen.[5]

## Specifications and access

| Property | Published specification |
| --- | --- |
| Stable Developer API model ID | `gemini-4-neon`[3] |
| Unified input and output window | 1,048,576 tokens[3] |
| Default synchronous output limit | 65,536 tokens[3] |
| Extended output | Up to the unified window using long decode, with input consuming part of that budget[6] |
| API thinking levels | `low`, `medium`, `high` (default), `xhigh`[4] |
| Knowledge cutoff | March 2026[2] |
| Cloud launch stage | Generally available, released November 18, 2026[7] |

The unified window includes the prompt, reasoning, tool calls, and final answer. A request containing a long input therefore has less room left for generation. The output allowance is not an additional million tokens on top of a million-token input.[6][10]

Availability depends on the product:

| Channel | Published access conditions |
| --- | --- |
| Gemini Developer API | Paid tier, through `generateContent` or the Interactions API[4] |
| Google AI Studio | Google AI Pro and Ultra plans, alongside API access[4] |
| [Gemini app](https://aiwiki.ai/wiki/gemini_app) | Google AI Pro and Ultra subscribers, through the model selector[1] |
| Antigravity | Paid Google AI Pro and Ultra subscribers across the app, CLI, and IDE extensions; enterprise access through Gemini for business[5] |
| Gemini Enterprise Agent Platform | Verified enterprise customers in selected regions; `global`, `us`, and `eu` endpoints[7] |

Antigravity's announced context limit is up to 512K tokens, separate from the API's 1,048,576-token window.[8] Product entitlements and context settings should not be inferred from the underlying model's maximum.

## Reasoning and long responses

The API's `thinking_level` controls the amount of reasoning. Google's Neon guide recommends `high` for complex tasks and `xhigh` when additional reasoning is worth greater latency and token consumption. `low` and `medium` reduce those costs for more responsive interactions.[4] These API defaults do not specify the default selected by every application.

Thinking uses the same generation allowance as the answer. Google's thinking documentation describes `thought` steps in the Interactions API, including an encrypted `signature` and an optional summary. The summary is a representation exposed by the API, rather than unrestricted access to the model's internal reasoning. `generateContent` represents signatures as metadata on response parts instead.[9]

Long decode splits a response into continuation segments. Neon permits up to 65,536 tokens per segment for streaming and non-streaming calls. With the Interactions API, setting `background=true` and a larger `max_output_tokens` lets Google's servers run the continuation loop; a client can instead manage continuation itself. Merely enabling background mode without increasing the limit still stops at 65,536 tokens.[6]

A minimal Interactions request can set the reasoning level explicitly:[4]

```python
from google import genai

client = genai.Client()
interaction = client.interactions.create(
    model="gemini-4-neon",
    input="Review this service design for race conditions.",
    generation_config={"thinking_level": "high"},
)
print(interaction.output_text)
```

This example uses the Google Gen AI SDK and its configured authentication. It makes a text request; access to a repository, terminal, or external service requires the surrounding application to provide those tools.

## Tools, files, and migration

The Developer API lists [function calling](https://aiwiki.ai/wiki/function_calling), [structured outputs](https://aiwiki.ai/wiki/structured_output), code execution, computer use, file search, URL context, and grounding with Google Search and Maps as supported. Agentic video understanding is supported but defaults to static processing. Tool retrieval is a preview capability. Neon does not provide image generation, audio generation, or the Live API.[3]

On Gemini Enterprise Agent Platform, Neon additionally accepts Office and OpenDocument formats, including Word, PowerPoint, Excel, ODT, ODP, and ODS files, inline or from Cloud Storage. The Cloud specification also lists system instructions and RAG Engine support, but not tuning.[7] These Cloud capabilities should not be assumed to have identical upload paths or limits in every consumer interface.

Migration from older models requires more than changing the model name. Google's Cloud developer guide identifies these API differences:[10]

| Integration concern | Neon behavior |
| --- | --- |
| Thinking configuration | Use `LOW`, `MEDIUM`, `HIGH`, or `XHIGH` on `GenerateContent`; `thinking_budget` and `MINIMAL` cause HTTP 400 errors |
| Sampling | Custom `temperature`, `top_p`, and `top_k` values are ignored |
| Parallel tool calls | Preserve thought signatures; only the first `functionCall` part in a parallel turn carries the signature |
| Conversation history | Preserve previous model turns and signatures rather than reconstructing their contents |
| Explicit cache objects | Unsupported; use implicit caching |
| Cloud location | Use `global`, `us`, or `eu` |

The guide also warns that continuation behavior depends on the API endpoint and SDK version. Its Cloud long-decode implementation uses `v1beta1`; changing the model, system instruction, or function declarations between continuation segments causes validation errors.[10]

## API prices and caching

The following Gemini Developer API prices were checked on November 22, 2026. Amounts are US dollars per million tokens; output prices include thinking tokens.[11]

| Inference mode | Uncached input | Cached input | Output |
| --- | --- | --- | --- |
| Standard | $3.00 | $0.15 | $15.00 |
| Batch | $1.50 | $0.075 | $7.50 |
| Flex | $1.50 | $0.075 | $7.50 |
| Priority | $5.40 | $0.27 | $27.00 |

There is no free API tier for Neon. Google's Neon guide specifies flat token pricing across the unified window.[4] The published output rate covers reasoning as well as visible answers; grounding can add separately billed search queries.[11]

[Context caching](https://aiwiki.ai/wiki/prompt_caching) is implicit and enabled by default, with a 4,096-token minimum for Neon. Repeated prompt prefixes can improve the chance of a hit, but meeting the threshold does not guarantee one. Google's caching documentation says Gemini 4 reprocesses the previous request's last 2,050 tokens at the standard input rate when the next turn adds fewer than 2,050 tokens. The response's `usage.total_cached_tokens` reports actual cached tokens.[12]

Token-price comparisons also need dates. Carbon's published Standard prices are temporarily $2 input and $10 output through November 30, 2026, rising to $4 and $20 on December 1. Consequently, Neon's $3/$15 is above Carbon's promotional rates and below its scheduled standard rates.[11]

## Evaluation results

Google published the following November 2026 results. They measure particular benchmarks and agent setups, rather than a single general measure of intelligence.[13]

| Benchmark | Neon result |
| --- | --- |
| CursorBench 4.1 | 65.4% |
| Terminal-Bench 4.1, cheating-adjusted accuracy | 65.2% |
| FrontierCode 2.0 | 46.2% |
| SWE-sweep, Antigravity provider harness | 39.0% |
| SWE-sweep, mini-SWE harness | 12.8% |
| Agents' Last Exam v1.1 | 75.8% |
| RiemannBench | 84.0% |
| Terminal-Bench-Science 0.1 | 70.5% |
| Chartography | 73.5% |
| GraphWalks, 256K-1M BFS F1 | 95.2% |
| CWE-bench v2 | 70.7% |
| SEC-bench Pro | 90.5% |

Google's methodology generally uses the highest thinking setting, except where stated, and reports pass@1 unless specified otherwise. FrontierCode uses `high`. The two SWE-sweep numbers use different harnesses; the Antigravity setup includes subagents, so the difference cannot be attributed solely to the model. Cybersecurity evaluations are an exception to the report's usual safety-filter setting.[13]

Independent evaluators provide further context. Artificial Analysis's November 18 analysis recorded an Intelligence Index score of 61 at `xhigh`. Its measured cost was $6.44 per task versus Carbon's $7.76 using non-discounted pricing, with approximately 21 versus 36 minutes per task. Those are measurements of that evaluation workload, not general API latency guarantees. Artificial Analysis also found that higher effort did not win every subtest: GDP.pdf scored 33.4% at `high` and 28.0% at `xhigh`.[14]

Zapier's AutomationBench leaderboard, version 1.0.6, listed Neon at 56.32% and $1.69 per task on `xhigh`, 54.34% and $0.88 on `high`, and 53.27% and $0.58 on `medium`. The benchmark grades final state changes in simulated business systems using deterministic assertions, rather than judging only an agent's written answer.[15]

Benchmark scores require their own interpretation. Cognition defines FrontierCode 2.0's score as a weighted rubric aggregate, assigning zero to solutions that fail blocking criteria.[16] Riemann-Bench's research paper describes privately held, expert-written mathematics problems with verifiable closed-form answers and agents permitted coding and search tools.[17] Neither score represents the fraction of arbitrary real-world jobs that Neon can complete.

## Safety and limitations

The model card acknowledges [hallucinations](https://aiwiki.ai/wiki/hallucination). Its March 2026 knowledge cutoff also means newer information requires external context or retrieval.[2] Google's announcement reports improvements in internal fabrication and agent-behavior tests, including a 99.2% pass rate on its agent honesty evaluation. Those are attributed evaluation results, not a guarantee of truthful answers or error-free tool use.[1]

For frontier risks, Google performed a Material Capability Change assessment against Gemini 4 Argon. It reported that Neon remained below the threshold requiring another full suite of critical-capability evaluations and carried forward Argon's threshold determinations. These included biological and cybersecurity critical-capability thresholds that could not be ruled out. The model card linked a forthcoming dedicated Safety & Security Report when checked on November 22, 2026.[2]

The separate cyber access described at launch was to roll out to trusted defenders through Google's Fairwind Program.[1] General availability of the standard model does not imply unrestricted availability of every cyber configuration or removal of usage policies.

## References

1. Google, Koray Kavukcuoglu. [Introducing Gemini 4 Neon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/). November 18, 2026.
2. Google DeepMind. [Gemini 4 Neon model card](https://deepmind.google/models/model-cards/gemini-4-neon/). November 18, 2026; accessed November 22, 2026.
3. Google. [Gemini 4 Neon, Gemini Developer API model specifications](https://aistudio.google.com/docs/models/gemini-4-neon). Accessed November 22, 2026.
4. Google. [What's new in Gemini 4 Neon](https://aistudio.google.com/docs/whats-new-gemini-4-neon). Accessed November 22, 2026.
5. Google Antigravity Team. [Gemini 4 Neon in Google Antigravity](https://antigravity.google/blog/gemini-4-neon-in-google-antigravity). November 18, 2026.
6. Google. [Long decode, Gemini Developer API](https://aistudio.google.com/docs/long-decode). Accessed November 22, 2026.
7. Google Cloud. [Gemini 4 Neon, Gemini Enterprise Agent Platform specifications](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-neon). Accessed November 22, 2026.
8. Google AI Developers Forum, rody. [Gemini 4 Neon and Orchestrator (Neon) are now available in Google Antigravity](https://discuss.ai.google.dev/t/gemini-4-neon-and-orchestrator-neon-are-now-available-in-google-antigravity/202301). November 18, 2026.
9. Google. [Gemini thinking](https://aistudio.google.com/docs/thinking). Accessed November 22, 2026.
10. Google Cloud. [Developer's guide to Gemini 4 Neon](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-4-neon). Accessed November 22, 2026.
11. Google. [Gemini Developer API pricing](https://aistudio.google.com/docs/pricing#gemini-4-neon). Accessed November 22, 2026.
12. Google. [Context caching](https://aistudio.google.com/docs/caching). Accessed November 22, 2026.
13. Google DeepMind. [Gemini 4 Neon model evaluation: Approach, methodology and results](https://storage.googleapis.com/deepmind-media/gemini/gemini_4_neon_model_evaluation.pdf). November 2026; accessed November 22, 2026.
14. Artificial Analysis. [Google continues to accelerate its release cadence with Gemini 4 Neon](https://artificialanalysis.ai/articles/gemini-4-neon-intelligence-gain/). November 18, 2026.
15. Zapier. [AutomationBench leaderboard](https://zapier.com/benchmarks). Version 1.0.6; accessed November 22, 2026.
16. Cognition. [FrontierCode 2.0](https://cognition.com/blog/frontier-code-2.0). October 27, 2026.
17. Suhaas Garre, Erik Knutsen, Sushant Mehta, and Edwin Chen. [Riemann-Bench: A Benchmark for Moonshot Mathematics](https://arxiv.org/abs/2604.06802). arXiv:2604.06802, version 3, June 23, 2026.

