Primary use stated or inferred cautiously from official documentation.
G
GOOGLE / gemini-4-argon
Gemini 4 Argon
Google model generally available since October 13, 2026. It accepts text, images, audio and video and generates text, including code; it does not generate images, audio or video. It uses a MoE architecture and a shared 1,048,576-token window. Robotics applications use this same model, not a separate endpoint.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | gemini-4-argon |
|---|---|
| Model type | Text and reasoning · Code · Multimodal · Robotics |
| Access model | Paid proprietary |
| License | Proprietary |
| Deployment | Hosted API |
| Release | 13 OCT 2026 |
| Knowledge cutoff | MAR 2026 |
| Entrance | Text · Image · Audio · Video |
| Exit | Text · Code |
| Context window | 1,048,576 tokens shared by input, reasoning and output |
| maximum output | 65,536 tokens by default; up to the remaining window with long decode |
| Reasoning | low, medium, high · medium by default |
| Published tools | Use of tools · Web search · Video understanding |
| Structured Outings | Not published |
| Batch processing | Sí |
| Prompt cache | Implicit caching; explicit caching is not supported |
| Fine-tuning | Not published |
| Verified platforms | Gemini API · Google AI Studio |
Software engineering, scientific and mathematical analysis, multimodal documents and multi-step agents. It offers three reasoning levels: low, medium and high, with medium as the default. Google publishes examples of spatial reasoning and robot control using tools and simulation. These are not safety guarantees or an autonomous controller ready to deploy without validation.
Input, thinking and response share one budget: a million tokens does not mean a million additional output tokens. Default output is 65,536 tokens; exceeding it with Interactions requires background execution and an explicit limit. Explicit caching is not supported, and temperature, top_p and top_k are ignored. The model can make mistakes and take several minutes. Its knowledge cutoff is March 2026.
03 / CAPABILITIES
What it can do
Orientation
Software engineering, scientific and mathematical analysis, multimodal documents and multi-step agents. It offers three reasoning levels: low, medium and high, with medium as the default. Google publishes examples of spatial reasoning and robot control using tools and simulation. These are not safety guarantees or an autonomous controller ready to deploy without validation.
Context
1,048,576 tokens shared by input, reasoning and output · 65,536 tokens by default; up to the remaining window with long decode
Tools and integration
Tool use · Web search · Video understanding
Access
Gemini API and Google AI Studio
Input modalities
TextImageAudioVideo
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | 2,00 USD / 1 M tokens | Standard API rate |
| Cached input | 0,10 USD / 1 M tokens | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | 10.00 USD / 1 M tokens | May include reasoning tokens |
| Batch input | 1.00 USD / 1 M tokens | Asynchronous processing |
| Batch output | 5.00 USD / 1 M tokens | Asynchronous processing |
Introductory USD prices per million tokens through November 30, 2026: input 2, output 10 including thinking, and implicit caching 0.10. From December 1 they become 4, 20 and 0.20. Batch and Flex cost half the standard rates; Priority has different prices. Searches incur separate charges after the shared allowance. There is no free tier or access through Google AI Pro. Numeric fields reflect standard and Batch rates in effect when verified.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Results published by Google, not reproduced by Inferama. We retain versions, subsets and conditions: GraphWalks uses F1 and two context ranges; OSWorld 2.0 is a partial score on the offline subset, taking the best of three runs; Chartography and LVBench are evaluated without tools. The methodology uses pass@1 and maximum reasoning unless otherwise stated. ERQA measures physical reasoning, not real-robot reliability. Compare costs, tools and budgets, not just percentages.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| Vals Index | 68,9 % | Score | View source ↗ |
| DeepSWE v1.1 | 77,9 % | pass@1 | View source ↗ |
| Terminal-Bench 4.0 | 57,4 % | pass@1 | View source ↗ |
| LABBench 2 | 88,8 % | Score | View source ↗ |
| GraphWalks (hasta 128K) | 99,7 % | F1 | View source ↗ |
| GraphWalks (256K–1M) | 84,2 % | F1 | View source ↗ |
| OSWorld 2.0 (offline) | 69,2 % | Partial score | View source ↗ |
| Chartography (no tools) | 71,6 % | Score | View source ↗ |
| LVBench | 91,7 % | Score | View source ↗ |
| CWE-bench v1 | 68,0 % | pass@1 | View source ↗ |
| ERQA | 87 % | Score | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Gemini 4 Argon
Version added to Inferama's verified catalog.
07 / SOURCES
Traceability
TECHNICAL SHEETFicha oficial del modeloai.google.dev↗
LAUNCHGeneral availability on October 13ai.google.dev↗
PRICESOfficial Standard and Batch pricingai.google.dev↗
MIGRATIONShared context and long decodeai.google.dev↗
MODEL CARDOfficial model card and resultsdeepmind.google↗
METHODOLOGYGemini 4 Argon evaluation methodologydeepmind.google↗
ROBOTICSPhysical reasoning and robotics applicationsdevelopers.googleblog.com↗