Gemini 4 Argon
Announced 2026-09-30 with staged access: Fairwind Program first, then paid API customers and Google AI Ultra 'as soon as possible'. Independent launch-day results: Artificial Analysis Intelligence Index 53 (tied with GPT-6 Astra), Arena Text #1 (1525). No public API model id, knowledge cutoff or model card as of 2026-09-30. Arena lists it as 'gemini-4-argon-high'. The 1M context comes from Artificial Analysis and Arena, not from Google. modality_in follows the Gemini family: AA's model page lists text + image, and its X post says text, image, video and speech. 'Argon' replaces the '3.x Pro' naming. Zero data retention is available for Fairwind partners using it as a managed model.
- Context window
- 1,000,000 tokens
- Max output
- 1,000,000 tokens
- Input
- text, image, audio, video, pdf
- Output
- text
- License
- proprietary
- Pricing
- input: $2 · output: $10 · cached input: $0.1 (per 1M tokens (introductory 50% discount, end date not announced; standard price $4 input / $20 output; cached input 95% off input)) source
- Verified
- 2026-09-30
How to call it
| Provider | Model id | Endpoint / URL | Docs |
|---|---|---|---|
| Google DeepMind Fairwind Program (vetted cyber defenders only; standalone or with the CodeMender agent) | — | deepmind.google/fairwind-program/ | — |
| Fairwind access form | — | rsvp.withgoogle.com/events/fairwind-program-interest-form | — |
| Gemini API / Google AI Ultra (announced as next, no date) | — | blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ | — |
Notable capabilities (6)
- FIRST 1M-token output limit: Can generate up to 1M output tokens in one response or trajectory (previous Gemini models: 64K). The Gemini API's new Long Decode Continuation feature pauses and resumes long responses across calls to avoid timeouts (per Artificial Analysis). source
- Enterprise knowledge work: Vendor-reported: Vals Index 68.9% (#1; confirmed on the Vals AI leaderboard), Harvey Legal Agent Benchmark 19.6% vs 6.7% for the next model, Vals Finance Agent v2 65.4%, Zapier AutomationBench 51.3% (#1). source
- Low hallucination rate: Artificial Analysis measured a 15% hallucination rate on AA-Omniscience, the lowest of any model scoring 45+ on its Intelligence Index (GPT-6 Astra 51%). source
- Frontier software engineering (mixed): Vendor-reported DeepSWE v1.1 77.9% (Opus 5.5 74.2%, GPT-6 Astra 74.1%), but it trails on FrontierSWE v2 (55.0% vs Astra 65.5%) and Terminal-Bench 4.0 (57.4% vs Opus 5.5 66.4%). Migrating C/C++ to Rust across Google, up to 800K+ lines. source
- Autonomous vulnerability discovery and patching: CWE-bench v1 68% pass@1, tied first with GPT-6 Astra and Grok 4.7 (public leaderboard). Offered without cyber guardrails to Fairwind defenders. Google reports 85.8% on its internal vulnerability benchmark and 70.9% on Wiz's pentest benchmark. source
- Long-context and long-video understanding: Vendor-reported GraphWalks BFS 256K–1M 84.2% (Astra 71.8%) and LVBench 91.7% (state of the art at launch). source
Google's first Gemini 4 model. It is not generally available yet, and there is no Gemini API or Vertex AI model id in the docs (checked 2026-09-30). Update this file with the model id, knowledge cutoff, docs link and model card once it reaches the Gemini API.
Benchmark highlights (Google's table at deepmind.google/models/gemini, with methodology in this PDF). Scores are for the highest thinking setting, and rival scores are mostly the providers' own figures.
| Benchmark | Argon | GPT-6 Astra | Fable 5.1 | Opus 5.5 |
|---|---|---|---|---|
| Vals Index | 68.9 | 63.1 | 65.8 | 67.0 |
| DeepSWE v1.1 | 77.9 | 74.1 | 67.4 | 74.2 |
| FrontierSWE v2 | 55.0 | 65.5 | 56.3 | 62.3 |
| Terminal-Bench 4.0 | 57.4 | 58.2 | 57.9 | 66.4 |
| Terminal-Bench Science 0.1 | 57.6 | 68.1 | 52.6 | 63.3 |
| GraphWalks 256K–1M | 84.2 | 71.8 | 65.0 | 66.8 |
| OSWorld-2.0 (offline, partial) | 69.2 | 72.6 | — | — |
| LVBench | 91.7 | 87.5 | 79.7 | 83.7 |
| CWE-bench v1 | 68.0 | 68.0 | 58.0 | 67.0 |
Timeline entry
- Google announces Gemini 4 Argon, its new frontier model, first released only to cyber defenders via the Fairwind Program ★★★★★
On Sept 30, 2026 Google DeepMind announced Gemini 4 Argon, its first new flagship since Gemini 3.1 Pro. It is a frontier model for coding, enterprise knowledge work and cyber defense, with a 1M-token output limit (previously 64K). Google's own table shows it leading or tied on 14 of 19 benchmark…
Other Google DeepMind models
Gemini 3.8 Flash TTS · Gemini 3.8 Live · Gemini 3.8 Flash · Gemini 3.5 Transcribe (and Transcribe Live) · Lyria 3.5 · Gemini 3.5 Flash-Lite · Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image) · Gemini Omni Flash (Omni 1.1 Flash) · Gemini Embedding 2 · Gemma 4 · Nano Banana 2 (Gemini 3.1 Flash Image) · Nano Banana Pro (Gemini 3 Pro Image) · Gemini Robotics 2 · Gemini Robotics ER 2 · Gemini Robotics On-Device 2 · Gemini 3.5 Live Translate · Gemini 3.1 Pro · Veo 3.1 · Genie 3 · Lyria RealTime · Gemini 3.7 Flash · Gemini 3.6 Flash · Gemini 3.5 Flash · Gemini 3.1 Flash TTS (preview) · Gemini 3.1 Flash Live (preview) · Lyria 3 (Clip / Pro) · Lyria 2 · Gemini 2.5 Flash Native Audio (Live, preview) · Gemini 2.5 Flash-Lite · Gemini 2.5 Flash · Gemini 2.5 Pro · Gemini 2.5 Flash TTS / Pro TTS · Gemini 3.1 Flash-Lite · Nano Banana (Gemini 2.5 Flash Image) · Gemini Robotics-ER 1.5 / 1.6 · Imagen 4