bloopcommitsGoogle DeepMind ne Gemini 4 Argon announce kiya, par zyaadatar log ise abhi use nahi kar sakte. Is guide mein seedhi baat hai: ye kya hai, kise mila, kya naya hai, benchmarks kya kehte hain, aur tumhe aaj kya karna chahiye.
Ek line mein
- Kaagaz pe strong, asli duniya mein abhi unproven.
- Access limited hai: sirf Google ke Fairwind Program ke cyber partners.
- Ye snapshot reel banate waqt ka hai. Numbers aur access badalte rehte hain, neeche ke official links dekho.
Kise mila, kise nahi
| Sawaal | Jawab |
|---|---|
| Abhi kaun use kar sakta hai? | Google ke Fairwind Program ke trusted cyber defenders. |
| Fairwind kya hai? | Limited-access program, 650 se zyada partners. Public apply link nahi hai. |
| Aam user ke liye? | Reel banate waqt Gemini app, API pricing page, Vertex AI ya AI Studio mein nahi tha. |
| Aage kise milega? | Google ke hisaab se agla number paid API customers aur Google AI Ultra ka hai. Koi date nahi di. |
| Free tier? | Nahi hai. |
Naya kya hai
Output limit
Output token limit 64K se badhkar 1 million tokens ho gayi. Input context window Google ne nahi bataya.
Cyber security
Google ka claim hai ki ye software ki critical kamzoriyan khud dhoondh sakta hai, check kar sakta hai aur patch kar sakta hai. Ye Google ka apna claim hai.
| Test (Google ka internal) | Argon | Gemini 3.8 Flash Cyber |
|---|---|---|
| Real-world vulnerability discovery | 85.8% | 71.0% |
| Wiz pentest benchmark | 70.9% | 58.2% |
Ye dono numbers Google ke apne datasets pe hain, yaani vendor-reported.
Benchmarks: Google ki apni table
Google ne 18 benchmarks ki table di. Usme sirf 4 models hain: Argon, GPT-6 Astra, Claude Opus 5.5 aur Claude Fable 5.1.
| Nateeja | Kitne |
|---|---|
| Argon jeeta | 12 |
| Barabar (tie) | 1 |
| Argon haara | 5 |
Jahan Argon jeeta (examples)
| Benchmark | Argon | Doosre |
|---|---|---|
| Vals Index v2.1 | 68.9 | Opus 5.5 67.0, Fable 5.1 65.8, Astra 63.1 |
| AutomationBench | 51.3 | Opus 5.5 42.5, Astra 41.4, Fable 5.1 31.4 |
Jahan Argon haara (examples)
| Benchmark | Argon | Jeetne wala |
|---|---|---|
| Terminal-Bench 4.0 | 57.4 | Opus 5.5 66.4 |
| FrontierSWE v2 | 55.0 | GPT-6 Astra 65.5 |
Reality check
- Google ki table mein Claude Sonnet 5.5 aur GPT-6.1 Sol hain hi nahi.
- Argon ke kai scores Google ne khud nikale hain, sabse zyada thinking setting pe.
- Independent boards ka nateeja mila-jula hai (neeche table).
- Access limited hai, isliye bahar walon ke hands-on test abhi nahi hain.
| Independent board | Argon ka haal |
|---|---|
| Vals Index | Number 1, 68.9% |
| Arena (text) | Number 1, lekin preliminary |
| Artificial Analysis Intelligence Index | Argon 52.6, Opus 5.5 57.6 (Argon peeche) |
Price (per 1 million tokens)
| Input | Output | |
|---|---|---|
| Shuruaati price | $2 | $10 |
| Baad mein | $4 | $20 |
Shuruaati price kab tak rahegi, ye Google ne nahi bataya.
Aaj tumhe kya karna chahiye
- Argon ka intezaar mat karo, access abhi limited hai.
- Jo models aaj khule hain unpe apna asli task chalao.
- Benchmark dekh ke nahi, apne kaam ke nateeje dekh ke model chuno.
- Kisi bhi "Argon free mein" wale link pe bharosa mat karo. Official page hi dekho.
Official sources
Google blog: Gemini 4 Argonblog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
Google DeepMind: Geminideepmind.google/models/gemini/
Fairwind Programblog.google/innovation-and-ai/technology/safety-security/fairwind-program/
Vals Index (independent)www.vals.ai/benchmarks/vals_index
Artificial Analysis (independent)artificialanalysis.ai/models/gemini-4-argon