Gemini 4 Argon Review: Google's New AI Model, Honestly Assessed
Google's Gemini 4 Argon is live worldwide. It leads on long documents, finance and coding, but loses several tests to rivals, and the price doubles on November 30.
Two weeks ago, Google's most powerful AI model was available to a handful of vetted security teams. Since October 13, anyone with an API key can use it. Gemini 4 Argon is now live, Google says it beats every rival on most of the benchmarks it chose to show, and the price is already scheduled to double. So is it as good as the headlines say?
What Gemini 4 Argon is
Argon is the first model in Google's Gemini 4 generation. According to its model card, it is a large sparse mixture-of-experts transformer built from scratch, not a fine-tune of an earlier Gemini. It accepts text, images, audio and video, but it only answers in text. Its knowledge stops at March 2026.
Key features
- A 1M-token window. Google says the output limit grew to 1 million tokens, up from 64,000, inside a single 1M unified input-and-output budget.
- Agentic work. Google highlights long, multi-step tasks in coding, finance, legal work, security and science, plus better spatial reasoning and visual coding.
- Cyber defence. Google says Argon can autonomously find, validate and patch serious software vulnerabilities. Security firm Wiz says it used the model to find a critical flaw in healthcare software that earlier models missed.
- Prompt-injection resistance. Google says it leads an independent injection test from Gray Swan, with attacks succeeding 0.7% of the time across 15 adaptive attempts.
How it scores
Google published a comparison against OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Opus 5.5. Here is a selection, with the winner in each row marked.
| Benchmark | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 | Claude Opus 5.5 | Leader |
|---|---|---|---|---|---|
| Vals Index | 68.9% | 63.1% | 65.8% | 67.0% | Argon |
| Vals Finance Agent v2 | 65.4% | 53.5% | 58.9% | 58.6% | Argon |
| DeepSWE v1.1 (coding) | 77.9% | 74.1% | 67.4% | 74.2% | Argon |
| GraphWalks 256k–1M (long context) | 84.2% | 71.8% | 65.0% | 66.8% | Argon |
| FrontierSWE v2 (coding) | 55.0% | 65.5% | 56.3% | 62.3% | GPT-6 Astra |
| Terminal-bench 4.0 | 57.4% | 58.2% | 57.9% | 66.4% | Opus 5.5 |
| Terminal-Bench Science 0.1 | 57.6% | 68.1% | 52.6% | 63.3% | GPT-6 Astra |
| OSWorld-2.0 (computer use) | 69.2% | 72.6% | n/a | n/a | GPT-6 Astra |
| CWE-bench v1 (security) | 68.0% | 68.0% | 58.0% | 67.0% | Tie |
Availability and pricing
Google's launch post says Argon is available worldwide from October 13, 2026.
| Where | Who gets it |
|---|---|
| Gemini API and Google AI Studio | Developers and businesses |
| Google Antigravity | Google AI Ultra subscribers |
| Gemini app and AI Mode in Search | Google AI Pro and Ultra subscribers |
| Fairwind Program | Approved cyber defenders, who also get a version without the cyber guardrails |
API pricing is $2 per million input tokens and $10 per million output tokens as an introductory rate, with cached input 95% cheaper. From November 30, 2026 it rises to $4 and $20. Google also says 10,000 scientists can apply for six months of free access to Antigravity's top usage tier.
Why it matters
The frontier is now a three-way race between Google, OpenAI and Anthropic, and no one model wins everything. Argon's edge is in the work businesses pay for: reading enormous documents, finance and legal agents, and fixing real code. Its giant output limit is also useful for tasks like migrating a whole codebase in one go.
The reality check
- The scores are mostly Google's own. The Vals Index is the main third-party number, and Vals lists Argon first at 68.9%, but that is a margin of under two points over Opus 5.5.
- It loses real tests. Terminal-bench 4.0, FrontierSWE, Terminal-Bench Science and OSWorld all go to a rival in Google's own table. "Best model" depends on your job.
- Benchmarks are not daily use. Speed, cost per finished task and how often it quietly gets things wrong only show up in real projects. The model card itself lists hallucinations and occasional slowness or timeouts.
- The safety picture is mixed. Google says it cannot rule out that Argon reaches its cyber and biological risk thresholds, and relies on safeguards to stay below them. Google's own report says it is still gathering feedback and tuning those guardrails.
- The price is temporary. Budget for $4 and $20, not $2 and $10.
- Text out only. No native image, audio or video generation from this model.
Our verdict
Argon looks like the strongest all-round model on paper, and especially good for long-document, finance, legal and general agent work. It is not a clean sweep, and until independent testers publish more results, treat the leaderboard claims as a strong marketing case rather than a settled verdict. If you build with AI, test it on your own tasks during the introductory pricing window before November 30.
Have you tried Gemini 4 Argon yet, and does it beat the model you use now?
Sources: Google: Introducing Gemini 4 Argon, Google: Gemini 4 Argon, our next era of frontier intelligence, Google DeepMind model card, TechCrunch, Vals AI. Benchmark figures are from Google's model card and have not been independently verified unless noted.
