THE SHORT VERSION
● Available now to paying customers. AI Pro in the Gemini app and AI Mode in Search; AI Ultra adds Antigravity and AI Studio; paid-tier API developers.
● $2 in / $10 out per million — until 30 November. After that it doubles to $4/$20. Cached input is 95% cheaper.
● State of the art on coding and long video. 77.9% on DeepSWE v1.1, 91.7% on LVBench, top of the Vals Index at 68.90%.
● Not a clean sweep. Around ninth on Artificial Analysis, and a 1,351-point Hacker News thread praised its prose while criticising its code.
Google announced Gemini 4 Argon on 30 September but shipped it only to a small group of cyber defenders through its Fairwind Program. That changed this week: Argon is now generally available to anyone paying Google.
Who gets it, where
| Who | Where Argon appears |
|---|
| Google AI Pro | Gemini app, AI Mode in Search |
| Google AI Ultra | Gemini app, AI Mode, Google Antigravity, AI Studio |
| Paid API developers | Gemini API paid tier, AI Studio, Android Studio |
| Enterprises | Gemini for business, API, Antigravity |
| Scientists and researchers | Free top usage tier in Antigravity for six months, 10,000 places, by application |
| Fairwind cyber defenders | A build without cyber guardrails, expanding to hundreds more partners |
| Free Gemini users | Not mentioned in either Google post |
Two details worth separating. The public version keeps its cyber guardrails; only vetted Fairwind partners get the build without them. And free users are simply absent from the announcement, which is usually how Google says "not yet" without saying it.
Pricing, and the deadline attached to it
| Period | Input / 1M | Output / 1M |
|---|
| Now to 30 November 2026 | $2 | $10 |
| From 1 December 2026 | $4 | $20 |
Cached input is priced 95% below the input rate, which matters more than it sounds for agent workloads that resend the same context repeatedly.
The 30 November date is the thing to plan around. If you are benchmarking Argon against alternatives on cost, benchmark it at $4/$20 — that is the price you will actually pay from December, and Vals AI already lists the post-introductory rate. Pricing a migration on the promotional number is how teams end up with a bill that doubles six weeks in.
Benchmarks
| Benchmark | Argon | Note |
|---|
| DeepSWE v1.1 | 77.9% | State of the art |
| LVBench, long video | 91.7% | State of the art |
| Vals Index | 68.90% | Claude Sonnet 5.5 67.04%, Claude Opus 5.5 66.97% |
| CWE-bench v1 | 68% | Tied for first, vulnerability remediation |
| AutomationBench (Zapier) | 51.3% | Ranked first |
| Artificial Analysis | ~9th | Composite index, different weighting |
That last row is the one to sit with. A model can hold state of the art on DeepSWE and the top of the Vals Index while placing around ninth on a different composite — because the indexes weight speed, cost and task mix differently. Neither is wrong. It means "best model" is not a question with one answer, and anyone quoting a single leaderboard is choosing the one that suits them.
Context limits
Google cited a 1M output token limit on 30 September. The general availability post describes a 1M unified input-output limit through the Gemini API and points to a continuation protocol for long outputs. Vals lists a 1M context window and 262k maximum output tokens.
Google has not said how much of that 1M budget a single response can consume, and it has not published rate limits for Pro against Ultra. If your use case depends on either, test it rather than planning from the headline figure.
What practitioners are saying
A Hacker News thread reaching 1,351 points divided fairly cleanly: strong praise for Argon's prose quality, pointed criticism of its code output. That is a familiar shape for a model topping a coding benchmark — benchmark tasks are well specified and self-contained, and real codebases are neither.
Worth weighing against the DeepSWE number rather than instead of it. Both are evidence.
Should you switch
● Long video or document understanding → test it now. 91.7% on LVBench is a real gap, not a rounding difference.
● Agent workloads that resend large context → the 95% cached-input discount may matter more to your bill than the headline rate.
● Production coding → run it on your own repository before believing either the benchmark or the Hacker News thread.
● Cost-sensitive at volume → model it at $4/$20 from December, not at the introductory price.
● On the free tier → nothing changes for you yet.
Sources
Google — Gemini 4 Argon ·
DealNTech — availability and pricing