> ## Documentation Index
> Fetch the complete documentation index at: https://docs.apiyi.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Gemini 4 Neon Is Live: The Middle Tier for Speed and Depth

> Google's Gemini 4 middle tier, GA on 18 November, is now on APIYI: four thinking levels, a 1M unified window and $3/$15 per 1M tokens. Google says it is about 3x faster than Carbon; Artificial Analysis independently measures an Intelligence Index of 61.

## Key Takeaways

* **Available now**: Google released Gemini 4 Neon on 18 November with same-day GA. It is listed on APIYI as `gemini-4-neon`, available in the `default` and `svip` groups on both Gemini-native and OpenAI-compatible endpoints
* **Priced at \$3 / \$15** per 1M tokens, with cached input at \$0.15. This is a **standard rate with no promotional expiry**
* **The price relationship flips on 1 December**: Carbon and Argon revert from their limited-time rates to \$4 / \$20, making Neon **25% cheaper** than both
* **The middle tier**: positioned between Carbon (hardest problems) and Hydrogen (everyday execution). Google says Neon is about 3x faster than Carbon and 32% cheaper per task at the `high` level
* **Four thinking levels, `high` by default**: `low`, `medium`, `high` and `xhigh`, plus a 1M unified window, up to 64K synchronous output tokens and long decode up to 1M
* **Independently benchmarked**: Artificial Analysis measures an Intelligence Index of 61 at `xhigh`, while also noting the model is "notably slow per task and very verbose" — details below

## Background

The Gemini 4 family shipped four models in about seven weeks: Argon was announced on 30 September and opened on 13 October; Carbon was announced on 28 October and opened on 10 November; Hydrogen filled the everyday execution slot on 12 November; Neon was announced on 18 November and reached **GA the same day** — the fastest rollout of the four.

Google does not position Neon as a cheaper Carbon or a stronger Hydrogen. It is the **middle solution for complex production workloads**: multi-file refactors, cross-system analysis, enterprise reporting and long-document synthesis, where reasoning depth matters but tasks also need to finish end to end. The launch materials claim roughly 3x the speed of Carbon, 32% lower cost per task at `high`, and an Intelligence Index only 1–2 points behind.

There is a pricing twist worth noticing. Carbon and Argon currently carry limited-time rates through 30 November 2026. Once they revert to standard pricing on 1 December, Neon's \$3 / \$15 becomes the **second most expensive model in the family** — only 25% cheaper than Carbon and Argon, not cheaper than them as it is today. Choose on the post-1 December price, not today's invoice.

## Detailed Analysis

### The family at a glance

| Model | Google's positioning | Thinking levels (default) | Current price (per 1M) | From 1 December |
| - | - | - | - | - |
| Gemini 4 Argon | Long-horizon coding, enterprise knowledge work, security | Three (`medium`) | \$2 / \$10 limited-time | \$4 / \$20 |
| Gemini 4 Carbon | Hardest reasoning, security audits | Three (`medium`) | \$2 / \$10 limited-time | \$4 / \$20 |
| Gemini 4 Neon | Complex coding, analysis, production work | Four (`high`) | \$3 / \$15 standard | Unchanged |
| Gemini 4 Hydrogen | Everyday execution, high-frequency agents | Four (`high`) | \$1.50 / \$7.50 standard | Unchanged |

Cached input is 95% off the input price across all four: \$0.15 for Neon, \$0.10 for Argon and Carbon, \$0.075 for Hydrogen.

### Official benchmarks

The following figures are **published by Google** and were not independently reproduced by APIYI:

| Benchmark | What it measures | Gemini 4 Neon |
| - | - | - |
| Artificial Analysis Intelligence Index | Composite intelligence | 61 (Google ranks it 3rd, `xhigh`) |
| Artificial Analysis Coding Agent Index | Coding agents | 58.6 (1st) |
| Terminal-Bench 4.1 | Terminal agent tasks | 65.2% |
| FrontierCode 2.0 | Front-end code (default `high`) | 46.2% |
| SWE-sweep | Software engineering (Antigravity / mini-SWE) | 39.0% / 12.8% |
| Zapier AutomationBench | End-to-end business workflows | 56.3% (1st) |
| Vals Index | Finance / coding / legal / tax composite | 70.9% (2nd) |
| Agent Arena | Real agent sessions (4,200+) | 1336 points (2nd) |
| HLE-Diamond | Hard reasoning | 71.3% |
| RiemannBench | Mathematical reasoning | 84.0% |
| Terminal-Bench Science | Scientific terminal tasks | 70.5% |
| Agents' Last Exam v1.1 | Long-horizon agent tasks | 75.8% |

Safety and behaviour numbers are also Google's: SEC-bench Pro 90.5%, CWE-bench v2 70.7% (3.4 points above Carbon), WebExploitBench v1.2 51.26%, AutoBaxBench hard tier 76%; a 0.0% cheating rate on Terminal-Bench 4.0, a 2.45% shortcut rate in the coding cheating evaluation, and a 0% attack success rate on Gray Swan's indirect prompt-injection benchmark at the highest thinking level. A separate "Gemini 4 Neon for Cyber" build is limited to verified defenders in the Fairwind Program.

Some results are highly sensitive to settings: FrontierCode uses the default `high` level, and SWE-sweep scores 12.8% on the mini-SWE harness but 39.0% inside Antigravity. Always cite the level and harness alongside the number.

### Independent view: same score, different rank

Artificial Analysis has listed Neon. Its independent measurements (retrieved 20 November 2026):

| Metric | `low` | `medium` | `high` | `xhigh` |
| - | - | - | - | - |
| Intelligence Index | 44 | 55 | 60 | 61 |
| Output speed | 97 tok/s | 101 tok/s | 101 tok/s | 110 tok/s |
| Time per task | 3.7 min | 8.3 min | 14.8 min | 21.5 min |
| Latency to first answer token | 13.9 s | 20.1 s | 40.6 s | 56.5 s |

Two things stand out:

1. **Same score, different rank.** Google calls 61 points 3rd overall; the AA leaderboard page (v5.0.1, 95 models) lists Neon 6th. The score agrees, the ranking methodology does not — cite the source and date whenever you quote a rank.
2. **AA is blunter than Google.** It describes Neon as "notably slow per task and very verbose." That is inherited from Carbon: Gemini 4 thinking tokens bill as output, so a model that thinks and writes more costs more. AA measures **\$6.44 per Intelligence Index task**.

**`xhigh` buys 1 index point over `high` while adding 16 seconds of first-answer latency and 6.7 minutes per task.** Unless that single point is exactly your business threshold, `high` is the better default.

### The four thinking levels

`thinkingLevel` is Neon's main cost and latency dial, and Google defaults it to `high`:

* `low`: high-throughput extraction, classification and short Q\&A
* `medium`: interactive coding, live UI, latency-sensitive tool calls
* `high` (default): complex coding, agent execution, multimodal document work
* `xhigh`: reserved for multi-step problems; AA's data shows a small gain for a large time cost

Thinking tokens are billed at the output rate. Google AI Studio's changelog notes that `temperature`, `top_p` and `top_k` were deprecated on 21 July 2026 — they do not error, but they are not guaranteed to take effect, so pipelines that relied on low temperature should move to prompt constraints and structured output instead.

### The 1M window and API behaviour

According to Google's model card and launch materials:

* **Context**: a 1M unified window shared by input, thinking and output
* **Output ceiling**: up to **64K tokens** per synchronous request; background long decode can extend to 1M
* **Multimodal input**: text, images, audio, video and documents
* **Knowledge cutoff**: **March 2026**, with some domains less current
* **Caching**: the model card does not provide Neon-specific rules; following the Gemini 4 family pattern, only implicit prefix caching is supported (minimum 4,096 tokens), not explicit `cachedContents`
* **Multi-turn and tool calls**: `thoughtSignature` must be returned verbatim or the request may fail with a 400 — a Gemini 4 family-wide requirement

On safety, the model card says Neon is broadly comparable to Hydrogen: text-to-text safety is unchanged (0.00pp), image-to-text safety improves by 1.90pp, and **unjustified refusals drop by 3.00pp**, meaning fewer cases where the model refuses a task it could handle. Its Frontier Safety MCC assessment is above Argon but below Google's threshold; biological and cybersecurity uplift, and loss-of-control risk, are both marked "cannot be ruled out." The full safety report is listed as "coming soon."

### APIYI availability

* Model name `gemini-4-neon`, available in the `default` and `svip` groups (per `enable_groups` in the pricing API)
* Native endpoint: `POST /v1beta/models/gemini-4-neon:generateContent`; OpenAI-compatible endpoint: `POST /v1/chat/completions`
* Billing: `model_ratio` 1.5, `completion_ratio` 5, `cache_ratio` 0.05 — that is, \$3 / \$15 with \$0.15 cached input, matching Google's list price line by line with no markup

> [!WARNING]
>
> No usable API test key was available on this machine when this article was written, so **we did not run pre-launch tests**. Everything above comes from Google's launch post, model card, AI Studio changelog and Artificial Analysis; on the platform side we only verified the pricing API. For a first run, cover both endpoints, streaming, `thinkingLevel` and tool calling before scaling up.

## Practical Usage

### Where it fits

* **Complex production coding**: multi-file refactors, code review, cross-module root-cause work — bolder than Hydrogen, faster than Carbon
* **Enterprise knowledge work**: synthesising long PDFs, spreadsheets and slide decks into cited reports — the Vals, Legal Research and Tax Agent benchmarks are aimed at this
* **The hard step inside a high-frequency agent loop**: run routine steps on Hydrogen and escalate planning steps to Neon
* **Assisted security review**: vulnerability discovery and fix suggestions (production defensive deployments should use the Cyber build with human review)

**Less suitable for**: simple extraction and classification (Hydrogen is half the price and faster), the truly hardest problems (Carbon has a higher ceiling), and latency-critical UI — AA measures more than 40 seconds to first answer token at `high`.

### Code examples

OpenAI-compatible endpoint:

```python theme={null}
from openai import OpenAI

client = OpenAI(
    api_key="sk-your-apiyi-key",
    base_url="https://api.apiyi.com/v1",
)

resp = client.chat.completions.create(
    model="gemini-4-neon",
    messages=[
        {"role": "system", "content": "You are a senior code reviewer. Report only reproducible issues."},
        {"role": "user", "content": "Review this concurrent code for data races and propose fixes: ..."},
    ],
    max_tokens=4096,
)
print(resp.choices[0].message.content)
```

Native endpoint with an explicit thinking level:

```bash theme={null}
curl https://api.apiyi.com/v1beta/models/gemini-4-neon:generateContent \
  -H "Authorization: Bearer $APIYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"role": "user", "parts": [{"text": "Find three anomalies in this earnings report and cite sources."}]}],
    "generationConfig": {
      "thinkingConfig": {"thinkingLevel": "high"}
    }
  }'
```

### Best practices

* **Start at `high`**: benchmark before dropping to `medium` or `low`; enable `xhigh` only when justified, since AA shows it adds both cost and time
* **Leave headroom in `max_tokens`**: thinking counts against the cap, and a small budget can return an empty body (Hydrogen testing showed 300 empties out while 600+ is stable)
* **Model cost without cache**: implicit caching hits sometimes but is not guaranteed; put stable prefixes first to improve hit rates
* **Stage long documents**: validate at 200K–300K tokens before pushing toward 1M
* **Return `thoughtSignature`**: multi-turn and tool-call chains are where it most often gets dropped
* **Stop tuning sampling parameters**: `temperature`, `top_p` and `top_k` are deprecated — constrain output with prompts and JSON Schema instead

## Pricing and availability

| Item | Price (per 1M tokens) |
| - | - |
| Input | \$3 |
| Output (including thinking) | \$15 |
| Cached input | \$0.15 (95% off input) |

* Model name `gemini-4-neon`, available in the `default` and `svip` groups.
* Native endpoint: `POST /v1beta/models/gemini-4-neon:generateContent`; OpenAI-compatible endpoint: `POST /v1/chat/completions`.
* One flat price across the full 1M window with no long-context surcharge; **a standard rate with no expiry date**.
* Recharge promotions can be stacked; see [Recharge Promotions](https://docs.apiyi.com/en/faq/recharge-promotions.md).

> [!NOTE]
>
> Model pricing is aligned with the official site and may change; the table above is for reference only. The **Model Pricing** tab in the top navigation is authoritative: [Model Pricing](https://docs.apiyi.com/en/models/index.md).

## Summary and recommendations

Neon fills the "depth and speed" slot. The real trade-off is: **about twice the price of Hydrogen for stronger reasoning, and a little less than Carbon for much faster end-to-end completion.**

A three-line rule:

* Everyday, high-volume, simple tasks → **Hydrogen** (\$1.50 / \$7.50, fast and cheap)
* Complex production tasks with clear acceptance criteria → **Neon** (default `high`, step down to `medium` to save)
* Open-ended hard problems, multi-day work, deep security analysis → **Carbon / Argon** (note the 30 November price increase)

For a first rollout, run your own eval set: the same tasks through Neon `high` and your current primary model, compared on quality × cost per task × first-answer latency. AA has already flagged verbosity, so if your workload is short-output and high-precision, constrain the format with JSON Schema and track what share of the bill is thinking tokens.

> [!NOTE]
>
> Official sources: `blog.google/intl/en-in/products/introducing-gemini-4-neon/`, `deepmind.google/models/model-cards/gemini-4-neon/` and `aistudio.google.com/docs/changelog` (18 November 2026 entry); independent evaluation: `artificialanalysis.ai/models/gemini-4-neon`; platform check: the APIYI pricing API at `api.apiyi.com/api/pricing`. Data retrieved 20 November 2026.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.