Google made the AI model Gemini 4 Neon generally available on November 18, 2026. Standard API pricing is $3 per million input tokens and $15 per million output tokens, twice the price of Hydrogen, which launched just before it. However, the introductory pricing for Carbon and Argon ends at the end of November, so Neon will be cheaper than both in December. To choose among the models Google has added for complex development and knowledge work, you need to compare more than the order of unit prices: you also need to consider how many tokens go to reasoning and how each model handles long documents.
Where does Neon fit?
The Gemini 4 models reached general availability in this order: Argon on October 13, Carbon on November 10, Hydrogen on November 12, and Neon on November 18. According to Google's API changelog, all four models arrived within 36 days, and three of them were added in November. Google lists complex software development, knowledge work, and science as Neon's target areas, and emphasizes its ability to complete tasks quickly from start to finish.
The models share some reasoning settings. Neon and Hydrogen let you choose how much to think before answering from four levels, low, medium, high, and xhigh, with high as the default. Carbon and Argon default to medium, with high as the maximum. Per Google's explanation of reasoning settings, any comparison that only switches the model name also picks up this difference in defaults. Being able to use xhigh is not unique to Neon.
The price gap reverses in December
Throughout November, Neon costs 50% more than Carbon and Argon for both input and output. But those two models will move from $2 input and $10 output through November 30 to $4 and $20 on December 1. For the same billed volume, Neon will then be 25% cheaper. The 2x gap with Hydrogen does not change. According to Google's pricing page, Neon is not eligible for the free tier.
For 100,000 input tokens and 10,000 billed output tokens, including reasoning, Neon costs $0.45 and Hydrogen costs $0.225. Including competitors, the cost of processing the same volume at standard API rates looks like this.
| Model | Input price ($) | Output price ($) | 100K input + 10K billed output ($) | Yen equivalent |
|---|---|---|---|---|
| Gemini 4 Hydrogen | 1.50 | 7.50 | 0.225 | approx. ¥36 |
| Gemini 4 Carbon / Argon (through Nov. 30) | 2 | 10 | 0.30 | approx. ¥48 |
| Gemini 4 Neon | 3 | 15 | 0.45 | approx. ¥72 |
| Gemini 4 Carbon / Argon (from Dec. 1) | 4 | 20 | 0.60 | approx. ¥96 |
| GPT-6.2 Sol | 2 | 10 | 0.30 | approx. ¥48 |
| GPT-6.2 Astra | 8 | 40 | 1.20 | approx. ¥191 |
| Claude Sonnet 5.5 | 2 | 10 | 0.30 | approx. ¥48 |
| Claude Opus 5.5 | 4 | 20 | 0.60 | approx. ¥96 |
| Claude Fable 5.2 | 8 | 40 | 1.20 | approx. ¥191 |
Unit prices are per million tokens for input and output respectively. This is a comparison as of November 19, 2026 (JST), calculated by the editorial team from the published standard prices of Google, OpenAI, and Anthropic. Caching, Batch, tool usage fees, and tax are not included. The yen conversion is a convenience calculation at $1 = ¥159.35, the midpoint of the Bank of Japan's 5 p.m. November 18 USD/JPY rate of ¥159.34–36, and differs from the rate at settlement.
Google's output price also covers the reasoning tokens used before an answer is produced. The "10K billed output" in the table is an assumed quantity that includes reasoning, not the length of the visible answer. The formula is input price × 0.1 + output price × 0.01. Because each company counts tokens differently and actual usage varies, this table compares the bill for the same volume. The cost of finishing the same job also depends on accuracy and the number of retries.
If you input the same 100,000 tokens, Neon's input alone costs $0.30, which exceeds Hydrogen's total of $0.225 in the example above. In other words, under these conditions, shorter Neon answers alone cannot bring the cost down to the same level. If you can halve both input and billed output to 50,000 and 5,000, Neon also comes to $0.225. How far you can cut input selection and rework after failures is the condition for closing the price gap.
For workloads that can tolerate waiting, Neon's Batch and Flex also offer a separate option at $1.50 input and $7.50 output. The results of a speed and quality comparison at standard prices cannot be applied directly to the cost of a different processing mode.
How a 1M-token window changes output design
The 1M-token window in Gemini 4 is a pool shared by input, reasoning, and output, and the normal response limit is 65,536 tokens. According to Google's long decode specification, the shared pool is exactly 1,048,576 tokens. The more input material you include, the less remains for reasoning and the answer.
For long decode, which generates long code or reports, you raise max_output_tokens and continue generation across multiple segments (hops). In background or streaming mode, all four models allow up to 65,536 tokens per hop. In non-streaming mode, Neon and Hydrogen allow 65,536, while Carbon and Argon allow 32,768. The normal response limit is the same across models, but for continued generation you need to choose the processing mode and the hop limit.
| Model | Context | Published output limit / method |
|---|---|---|
| The four Gemini 4 models | 1,048,576 shared by input, reasoning, and output | Normally 65,536 for reasoning + output. Long decode continues up to the shared pool |
| GPT-6.2 Sol / Astra | 1,050,000, max input 922,000 | Max output 128,000 |
| Claude Sonnet 5.5 / Opus 5.5 / Fable 5.2 | 1M | Max output 128K |
Sources: Google's long decode specification, OpenAI's Sol specification and Astra specification, and Anthropic's model overview. M and K follow each company's notation. Capacity is not a measure of accuracy or generation speed.
This difference matters depending on whether you want to receive a long deliverable in a single response or can build continued processing into your workflow. Both GPT-6.2 models also have a condition under which, when input exceeds 272K, the input and cache unit prices for the entire request double and the output unit price rises 1.5x. Because the pricing conditions differ from the 100,000-token input in the previous section, designs that pass in large volumes of material at once need to be recalculated.
Consider, for example, a job that examines formula inconsistencies across sales workbooks from multiple departments and produces proposed fixes and explanations. The Gemini API update the same day made it possible to specify relevant sheets and cell ranges in Office and OpenDocument files. According to the document processing API documentation, for spreadsheets you can select a sheet number and an A1-style cell range, and also specify whether to read data, formulas, or visual. Because only the targeted portion is billed, a design that first examines the formulas in the relevant sheets and adds surrounding cells or visuals as needed can reduce the volume of rereading entire workbooks each time.
Document selection is an API feature available in Gemini 3 and Gemini 4 and later. To measure the effect of choosing Neon, pass the same range to Hydrogen too, and compare whether it misses formula errors and whether its proposed fixes are usable. This lets you separate the effect of the model price difference from the effect of improving how input is supplied.
What existing models' results say about using each one
Neon's own comparison scores, measured completion times, and cost per task could not be confirmed in the official materials reviewed this time. However, the evaluations of existing models published by Google DeepMind show on which tasks candidates should be compared. The table below extracts three metrics from the Carbon and Argon comparison table, covering software development, graph traversal in long text, and mathematics.
| Evaluation / condition | Carbon | Argon | GPT-6 Astra | Claude Opus 5.5 | Claude Fable 5.1 |
|---|---|---|---|---|---|
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| GraphWalks (256k–1M, BFS F1) | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| RiemannBench | 72.0% | 76.0% | 72.0% | 69.6% | 65.6% |
Carbon posts a high value on GraphWalks, but Argon beats it on RiemannBench. On FrontierSWE 2, GPT-6 Astra is higher. The rankings are not the same for tasks that trace relationships in long information, for mathematics, and for code fixes.
For Hydrogen, a separate table is published with different comparison targets and evaluation versions.
| Evaluation / condition | Hydrogen | Claude Sonnet 5.5 | GPT-6.1 Terra |
|---|---|---|---|
| SWE-sweep | 13.5% | 3.1% | 5.8% |
| Terminal-Bench 4.1 (cheating-adjusted) | 57.0% | 63.9% | — |
| OSWorld-2.1 (offline subset) | 77.6% | 83.9% | 74.2% |
Hydrogen leads on SWE-sweep, but Sonnet is higher on Terminal-Bench, which evaluates terminal work, and on OSWorld, which evaluates screen operation. This is a reason not to narrow candidates by low unit price alone, and instead to choose which models to try starting from the evaluations closest to your own work.
Both tables use values published by Google and are not remeasured by the editorial team. GPT-6 Astra and Fable 5.1 in the Carbon and Argon table are a different generation from GPT-6.2 Astra and Fable 5.2 in the pricing table. They cannot be read as an overall ranking stitched together from Hydrogen's separate evaluations or different versions, nor as price-performance for the latest generation.
What to compare first before adopting
If you try Neon, it is best to start by giving it, under the same conditions, the code fixes Hydrogen could not pass and the judgment tasks that draw on multiple documents. Google recommends evaluating each reasoning level and choosing the lowest setting that meets the required quality. xhigh increases waiting time and token usage, and for simple extraction the reasoning can be excessive. Trying the default high once makes it hard to tell whether the model or the setting made the difference.
In the sales workbook example above, you would pass the same cell range and fix the pass criteria for the errors to detect and the proposed fixes. Then record total input and total output including reasoning, and compare the number of retries and the time to pass. For tasks where Neon passes in one attempt and reduces Hydrogen's retries, even double the unit price is worth considering. For extraction jobs that both pass, there remains a basis for choosing Hydrogen's lower price.
There are also alternatives to consolidating everything on one model. According to Google's Hydrogen announcement, Antigravity Orchestrator dynamically switches between models such as Carbon and Hydrogen within the same session. On Terminal-Bench 4.0 it kept a task success rate close to Carbon's, with 24% lower average quota consumption and 19% faster completion. This is Google's measurement of switching between those two models and does not show Neon's speed difference on its own, but it is material for evaluating an operation that hands only the difficult parts to the more expensive model.
On December 1, the pricing terms for Carbon and Argon change. Rather than fixing your pre-adoption comparison as is, remeasure how far Neon reduces the processing needed to reach a pass, so you can separate the work worth paying extra for from the work to leave to a lower-priced model.
