On October 13, 2026, Google opened Gemini 4 Argon to all developers, but some features don't work: there is no free tier, no explicit caching, no Live API and no image generation, and the public version has its cyber capabilities restricted. Since the September 30 announcement, only Fairwind program members and invited testers had access (we covered that announcement earlier). The model is already on GPTunneL, with a 50% discount right now.
What doesn't work in Argon
Restrictions from the Gemini API documentation and Google's announcement:
| What doesn't work | What it means in practice |
|---|---|
| Free tier | you can't try it without paying, neither in AI Studio nor in the API |
| Explicit caching | implicit only: Google decides what to cache, you can't pin a long context in advance |
| Live API | no real-time voice or video dialogue |
| Image and audio generation | text output only |
| Long output out of the box | by default the response is capped at 65,536 tokens; a separate long decode mode stretches it to 1,048,576 |
| Full cyber capabilities | the public version refuses harmful requests; only Fairwind program members get Argon without cyber restrictions |
The context window is 1,048,576 tokens, shared between input and output. Thinking levels are low, medium and high.
Who can use it and where
The model ID in the API is gemini-4-argon. Google calls the access global, and there are several channels:
- developers and companies - the paid tier of the Gemini API in Google AI Studio and Android Studio;
- Google AI Ultra subscribers - AI Studio and the Antigravity platform;
- Google AI Pro and Ultra subscribers - the Gemini app and AI Mode in Search;
- cyber defenders - the Fairwind program, which is expanding to hundreds of partners: they get Argon without cyber restrictions, and applications are reviewed on a rolling basis.
How much Gemini 4 Argon costs
Argon is currently 50% off on GPTunneL. For comparison, here are prices of neighboring models on our price list, per 1K tokens:
| Model | Input | Output | Request: 100K input, 10K output |
|---|---|---|---|
| Gemini 4 Argon, 50% off | $0.006 ($0.012 without discount) | $0.03 ($0.06) | $0.90 ($1.80) |
| Claude Opus 5.5 | $0.012 | $0.06 | $1.80 |
| Claude Sonnet 5.5 | $0.004 | $0.02 | $0.60 |
| GPT 6 Astra | $0.02 | $0.10 | $3.00 |
| Gemini 3.8 Flash, 50% off | $0.0015 | $0.0075 | $0.225 |
The last column is a typical request: a large 100,000-token document and a detailed 10,000-token answer. Without the discount Argon costs the same as Opus 5.5; with it, half as much. Current prices are on the pricing page.
What makes Argon different
Google built the model for long work with a verifiable result. The examples below come from the September 30 announcement; they are the company's internal data without independent verification.
Long-running development
- Porting C/C++ to Rust: from libraries of tens of thousands of lines to the Zircon kernel in Fuchsia with 800,000+ lines, with tests, emulation and review.
- The libgav1 video decoder became 2.7 times faster than the previous Rust port: agents replaced 32,000 lines of SIMD code, and the output video is identical.
- Data center memory: using profiling telemetry, agents freed up more than 300 TiB; Google estimates total savings at 500 TiB - 1 PiB.
Cyber defense
The main focus of the announcement. Argon finds critical vulnerabilities on its own, confirms them and prepares fixes. With it, Wiz found a critical vulnerability in medical software used by hospitals worldwide: it exposed personal data, and previous frontier models had missed it. On CWE-bench v1 (fixing vulnerabilities) Argon scores 68% - tied for first place.
That is why the model was released in stages: what finds vulnerabilities also helps attackers. Only vetted defenders in the Fairwind program get it without restrictions - government bodies, healthcare, telecom. The program is expanding to hundreds of partners, and its rules forbid reselling that access.
Finance and law
Google claims the lead on Vals Finance Agent v2 (multi-step financial research) and Harvey's Legal Agent Benchmark (legal research and documents), but gives no numbers for them. A high score here doesn't replace verification: frame the task so that every conclusion cites a document and data gaps are listed separately.
Robots and attack protection
- Robotics: 87% on ERQA (reasoning about the physical world) - according to Google, the best result among frontier models. A separate guide for robotics was released.
- Prompt injection: on Gray Swan IPI, 0.7% of attacks via instructions hidden in data succeed over 15 adaptive attempts. The agent's reasoning is monitored, and it is stopped if it goes beyond the scope of the task.
What the benchmarks show
Independent rankings: Argon leads in two and trails in two others. Comparison with its closest rivals:
| Test | Argon | Opus 5.5 | Sonnet 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| Vals Index 2.1, score (Oct 13) | 68.90% (1st) | 66.97% (3rd) | 67.04% (2nd) | 63.13% (7th) |
| Artificial Analysis Intelligence Index (Oct 1) | 53 | 58 | 56 | 53 |
| Terminal Bench 4, command-line agent (Oct 1) | 57% | 60% | 64% | 59% |
What follows from this:
- Argon is first on the Vals Index and finishes the test faster than its rivals: 46 minutes versus 1 hour 19 for Opus 5.5. But the same ranking puts Sonnet 5.5 above Opus 5.5, and The Decoder advises reading this result with caution.
- Half of the Vals Index is about finance: tasks are weighted by each industry's share of US GDP, finance makes up about 52% of the index and software development about 37%. In another field the gap may differ. For scale: Gemini 3.8 Flash scored 54.83% there as of September 30.
- On the Artificial Analysis index, Argon is level with GPT-6 Astra and 5 points behind Opus 5.5. On Terminal Bench 4 it trails all three rivals, including Sonnet 5.5.
- It hallucinates less but is less accurate. On AA-Omniscience, Argon makes up an answer 15% of the time versus 51% for GPT-6 Astra, but gives correct answers 50% of the time versus 63%.
- Argon uses a lot of tokens. In Artificial Analysis tests it produces an average of 62,000 output tokens per task versus 27,000 for GPT-6 Astra, 2.3 times more. You pay for tokens, so compare models by the cost of a task, not the price per 1K.
- In Text Arena (human ratings in blind comparison) Argon ranks first, with 1525 points, 20 ahead of Claude Opus 4.6. In Code Arena (WebDev) it is eighth: 1679.
Artificial Analysis and Arena results are taken from The Decoder review of October 1; Vals Index scores are from the Vals AI table as of October 13.
Google's own benchmarks
These numbers come from Google itself; percentages from different tests can't be compared with each other:
| Test | What it measures | Argon |
|---|---|---|
| DeepSWE v1.1 | long-running development tasks | 77.9% |
| CWE-bench v1 | fixing vulnerabilities | 68% |
| AutomationBench | end-to-end business processes | 51.3% |
| LVBench | understanding long videos | 91.7% |
| ERQA | reasoning about the physical world | 87% |
| Gray Swan IPI | share of successful attacks via data, lower is better | 0.7% |
What this means for you
If you want to try it. With the discount, Argon is half the price of Opus 5.5 with comparable benchmark results. But plan a budget for ongoing work on the undiscounted price: $0.012 and $0.06 per 1K tokens, the same as Opus.
If you have a large repeated context (the same knowledge base in every request). Argon has no explicit caching, so check on a real task what resending the context costs.
If you want to know whether you need Argon. Collect 10-20 of your own long tasks with acceptance criteria (code migration, vulnerability analysis, a financial model) and run them on your current model and on Argon: share of accepted results, number of manual fixes, time and cost. Short tasks your current model already handles don't need to move.
If you need a terminal agent or coding. On Terminal Bench 4 Argon is weaker than Sonnet 5.5 and Opus 5.5 and uses more tokens. Sonnet 5.5 is also three times cheaper than Argon without the discount.
If your tasks are simple (classification, translation, short answers). Argon is overkill for them: Gemini 3.8 Flash currently costs four times less.
Where to try Gemini 4 Argon on GPTunneL
Gemini 4 Argon is already in the GPTunneL catalog; in the API it is called gemini-4-argon, with a 1,048,576-token context. With the 50% discount it costs $0.006 per 1K input tokens and $0.03 per 1K output tokens. It's easy to compare it with Opus 5.5, Sonnet 5.5 or Gemini 3.8 Flash on your own task in one chat. No VPN needed, local payment methods, no subscription, one shared balance.
Sources
- Gemini 4 Argon: our next era of frontier intelligence, Google, 30.09.2026 - announcement: examples of work inside Google, the Wiz case, Google's own benchmarks.
- Gemini 4 Argon: Built to solve your hardest problems, Google, 13.10.2026 - availability, benchmarks.
- Build with Gemini 4 Argon, Google, 13.10.2026 - developer launch.
- Gemini 4 Argon in the Gemini API documentation, Google, 13.10.2026 - limits, thinking modes, unsupported features.
- Vals Index, Vals AI, 13.10.2026 - scores and duration of runs.
- Google Gemini 4 Argon closes the gap with OpenAI and Anthropic but doesn't take a clear lead, The Decoder, 01.10.2026 - Artificial Analysis and Arena results.