Last updated:
Gemini 4 Argon: price, access and benchmarks
Gemini 4 Argon is Google DeepMind's new frontier model and the first of the Gemini 4 generation. Google announced it on September 30, 2026, but only a small group of security partners can use it so far.
Key facts
| Maker | Google DeepMind |
|---|---|
| Announced | September 30, 2026 |
| Who can use it now | Security teams in Google's Fairwind Program |
| Next to get it | Paid Gemini API customers and Google AI Ultra subscribers, "as soon as possible" per Google (no date) |
| Launch price per 1M | $2 input, $10 output, $0.10 cached input |
| Standard price per 1M | $4 input, $20 output, after the launch period |
| Max output | 1M tokens per response (Gemini models were capped at 64K before) |
| Strong at | Agentic coding, legal and finance work, long context, charts and long video, finding and patching security bugs |
Source: Google: Gemini 4 Argon announcement
Gemini 4 Argon API pricing
| Price | Input | Cached input | Output |
|---|---|---|---|
| Launch price | $2 | $0.10 | $10 |
| Standard price | $4 | $0.20 | $20 |
- Cached input is 95% off the input price.
- Google has not said when the launch price ends. After it, Argon costs $4 / $20, the same list price as Claude Opus 5.5.
- GPT-6 Astra costs $10 / $50, so Argon is cheaper than OpenAI's flagship even at its standard price.
Source: Google: Gemini 4 Argon announcement
How to get Gemini 4 Argon
- Today: only through the Fairwind Program, Google's scheme for vetted cyber defenders such as governments, critical infrastructure operators and security labs. Members get a version without the cyber guardrails, for defensive work only.
- Next: paid Gemini API customers and Google AI Ultra subscribers. Google says this comes "as soon as possible" and has not given a date.
- Why the wait: Argon is trained to find and patch software vulnerabilities on its own, and Google is taking part in the US government's voluntary pre-release testing before a wider launch.
- Until then: if you need a frontier model today at Argon's standard price, Claude Opus 5.5 is on general release at the same $4 / $20.
Benchmarks (Google's numbers)
Google's table from the Gemini model page. Google ran every model at its highest thinking setting. The best score in each row is in bold.
| Benchmark | Gemini 4 Argon | Opus 5.5 | GPT-6 Astra | Fable 5.1 |
|---|---|---|---|---|
| Vals Index | 68.9% | 67.0% | 63.1% | 65.8% |
| AutomationBench | 51.3% | 42.5% | 41.4% | 31.4% |
| Vals Finance Agent v2 | 65.4% | 58.6% | 53.5% | 58.9% |
| Harvey's Legal Agent Benchmark | 19.6% | 3.8% | 5.4% | 6.7% |
| DeepSWE v1.1 | 77.9% | 74.2% | 74.1% | 67.4% |
| FrontierSWE v2 | 55.0% | 62.3% | 65.5% | 56.3% |
| Vibe Code Bench | 91.9% | 90.3% | 89.6% | 90.3% |
| Terminal-Bench 4.0 | 57.4% | 66.4% | 58.2% | 57.9% |
| PostTrainBench | 45.3% | 49.3% | 44.3% | 40.2% |
| Terminal-Bench Science 0.1 | 57.6% | 63.3% | 68.1% | 52.6% |
| LABBench 2 | 88.8% | 73.1% | 85.4% | 68.6% |
| RiemannBench | 76.0% | 69.6% | 72.0% | 65.6% |
| GraphWalks BFS (up to 128K) | 99.7% | 90.6% | 98.7% | 91.4% |
| GraphWalks BFS (256K-1M) | 84.2% | 66.8% | 71.8% | 65.0% |
| Agent's Last Exam | 39.5% | 38.2% | 34.2% | — |
| OSWorld-2.0 | 69.2% | — | 72.6% | — |
| Chartography | 71.6% | 66.3% | 71.0% | 46.2% |
| LVBench | 91.7% | 83.7% | 87.5% | 79.7% |
| CWE-bench v1 | 68.0% | 67.0% | 68.0% | 58.0% |
Biggest leads: Harvey's legal agent benchmark (19.6% vs 3.8% for Opus 5.5), AutomationBench (51.3% vs 42.5%), long-context GraphWalks from 256K to 1M (84.2% vs 66.8%) and DeepSWE (77.9% vs 74.2%).
Where it trails: Terminal-Bench 4.0 (57.4% vs 66.4% for Opus 5.5), FrontierSWE v2 (55.0% vs 65.5% for GPT-6 Astra), Terminal-Bench Science and OSWorld-2.0.
Source: Google DeepMind: Gemini model page
Source: Google DeepMind: evaluation methodology (PDF)
Cost per coding task
List prices with our calculator's assumptions, per agent task:
| Codebase size | Argon (launch) | Argon (standard) | Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| Small (< 10k lines) | $0.112 | $0.223 | $0.223 | $0.576 |
| Medium (10k-100k lines) | $0.249 | $0.498 | $0.498 | $1.290 |
| Large / monorepo (> 100k lines) | $0.594 | $1.188 | $1.188 | $3.090 |
At 173 tasks a month on a medium codebase: about $43 on Argon at the launch price ($86 at the standard price), $86 on Claude Opus 5.5 and $223 on GPT-6 Astra. These use the same token counts for every model; real usage differs by model.
Gemini 4 Argon vs Opus 5.5 Gemini 4 Argon vs GPT-6 Astra All model prices
What the press says
- VentureBeat: Says Google retook the benchmark lead from OpenAI and Anthropic, but only in a limited release.
- Axios: Google's first new flagship in almost a year, after Gemini 3.5 Pro never shipped.
- 9to5Google: Covers the launch and standard prices and the roll-out order.
- MarkTechPost: Focuses on the 1M-token output limit and the cyber-defense use case.
FAQ
Is Gemini 4 Argon out?
It was announced on September 30, 2026, but only members of Google's Fairwind cyber-defense program can use it. Paid API customers and Google AI Ultra subscribers are next, with no date given.
How much does Gemini 4 Argon cost?
$2 per 1M input tokens and $10 per 1M output tokens at launch, with cached input at $0.10. After the launch period it goes to $4 / $20.
Can I use Gemini 4 Argon for free?
No. Google has only named paid routes: the paid Gemini API and the Google AI Ultra subscription, and neither is open yet.
Is Gemini 4 Argon better than Claude Opus 5.5?
On Google's table it beats Opus 5.5 on most benchmarks, including DeepSWE (77.9% vs 74.2%). Opus 5.5 is ahead on Terminal-Bench 4.0 (66.4% vs 57.4%). See our full comparison.
What is the Fairwind Program?
Google's early-access scheme for trusted cyber defenders, such as governments, critical infrastructure and security labs. They get Argon without its cyber guardrails for defensive security work.
Independent page. Not affiliated with Google. Benchmarks are Google's own; we have not run our own tests.