Google has put a clock on its cheapest frontier AI pricing. In an October 13 announcement, the company said Gemini 4 Argon is now generally available to developers and businesses worldwide, and for the first time it attached an end date to the introductory API price: $2 per million input tokens and $10 per million output tokens is good only through November 30, 2026. From December 1, the standard price doubles to $4 and $20.
Two weeks ago, Argon was a limited release. Google unveiled it on September 30 with strong benchmark claims but kept public access restricted to cybersecurity partners in its Fairwind Program, and the $2/$10 price came with no stated expiry. That ambiguity is now gone. The October 13 post, authored by Google DeepMind’s Koray Kavukcuoglu, says Argon is “now globally available today” and puts the price doubling in a footnote, which is the kind of detail enterprise buyers plan budgets around.
Where Argon is available now
The model ID is gemini-4-argon, and it now runs in the paid Gemini API, Google AI Studio, and Google’s Antigravity agentic coding environment (where individual access needs the Ultra plan). Google AI Pro and Ultra subscribers get it in the Gemini app and in AI Mode in Search. Android developers can call it from Android Studio (Canary) and Firebase AI Logic, and enterprise teams through Gemini Enterprise and Gemini for business.
What it is not on: the API’s free tier. Developers on the free tier will need a paid plan to touch Argon at all. Google is also widening its Fairwind Program to hundreds of additional security partners, and vetted cyber defenders who apply can get a version of Argon without the cyber guardrails, for defensive work. Separately, Google is offering 10,000 scientists its top Antigravity usage tier free for six months.
The price table that matters: $2/$10 now, $4/$20 in December
Here is how Argon’s pricing stacks up against the current frontier field, using VentureBeat’s comparison (per million tokens):
| Model | Input / 1M | Output / 1M |
|---|---|---|
| Gemini 4 Argon, through Nov 30 | $2 | $10 |
| Claude Sonnet 5.5 | $2 | $10 |
| GPT-6.1 Sol | $2 | $10 |
| Gemini 4 Argon, from Dec 1 | $4 | $20 |
| Claude Opus 5.5 | $4 | $20 |
| Meta Muse Wonder | $4 | $20 |
| GPT-6 Astra | $10 | $50 |
At today’s exchange rate of roughly 87.5 rupees to the dollar, the introductory price works out to about Rs 175 per million input tokens and Rs 875 per million output tokens, rising to roughly Rs 350 and Rs 1,750 in December. Cached input, which Google discounts 95%, costs about $0.10 per million tokens during the intro period and $0.20 after. Budget-conscious teams also get cheaper tiers: Batch and Flex run $1/$5 in the intro window, and Priority costs $3.60/$18.
For the fuller picture across every major lab, see our October 2026 AI price list. The December doubling moves Argon from the economy tier (alongside Claude Sonnet 5.5 and GPT-6.1 Sol) into the premium tier occupied by Meta’s Muse Wonder and Claude Opus 5.5. If you are choosing between them, the next six weeks are the cheapest window to run a real head-to-head.
The one-million-token catch: input and output share a budget
Argon has an unusual limit that changes how you budget long agentic runs. Google restates a 1-million unified token budget, meaning input and output draw from the same pool. A 600,000-token prompt leaves only 400,000 tokens for the response, which is a different constraint than rivals that cap context and output separately. Google points developers to a new “continuation protocol” in the docs for sustained multi-step work.
Caching helps, but with a wrinkle: Google offers automatic implicit prompt caching rather than letting developers explicitly create caches through the API. When cached content is reused successfully, eligible input gets the 95% discount. For applications with large recurring contexts, that automatic behavior is convenient but gives you less control than explicit cache management.
Google’s claims, and what is still unverified
The October 13 post adds fresh performance claims to the launch story: number one on the Vals Index, a 35%-plus increase in agentic code submissions and nearly one-third fewer rollbacks for Google teams using Argon in Antigravity, a 2.85x inference speedup for Gemma 4 E2B on an A100 via vLLM, and a 0.7% indirect prompt-injection attack success rate on Gray Swan’s tests across 15 adaptive attempts. Early enterprise pilots named include Amplitude, Deloitte, Rakuten, LexisNexis, and Ambience Healthcare.
Two caveats. First, every one of these numbers comes from Google; independent benchmarks are still catching up, and at least one aggregation of third-party testing describes Argon’s comparative standing as mixed. Second, the same week Google’s agent story made headlines: its new Gemini agent is being pitched as a colleague-class worker. Argon is the model most likely to power that push, which makes the December price a number worth watching for anyone building agents on Google’s stack.
What to do before November 30
If you run production workloads on a frontier model, the play is straightforward: use the next six weeks as a subsidized evaluation window. Port a representative slice of traffic to gemini-4-argon, measure cost per completed task (not per token, since reasoning-heavy models can burn through a cheap rate), and decide whether the December price still wins against Claude Sonnet 5.5 or GPT-6.1 Sol at their flat $2/$10. Build your caching assumptions around implicit caching, and design long agent loops around the unified 1M budget with the continuation protocol from day one.
The intro price was never going to last. Now there is a date on it, and December 1 is 47 days away.
