Google's frontier model is finally on sale, and its real price is dated 1 December
For six weeks Google had a frontier model it could not sell you.
Gemini 4 Argon was announced on 30 September 2026 with a strong set of first-party numbers — 77.9% on DeepSWE v1.1, 51.3% on AutomationBench for a claimed first place, 91.7% on LVBench, 68% on CWE-bench v1 — and then handed to “trusted cyber defenders” through Google’s gated Fairwind programme. It appeared on no public rate card. It had no date. Google’s position at the top of the capability curve was an announcement rather than a product, which is why our Gemini review has carried the line no frontier model on sale since the start of October.
On 14 October 2026 that ended. Google Cloud’s announcement says Argon is generally available “starting today”, with the model ID gemini-4-argon, across the Gemini platform API, Google AI Studio, Google Antigravity, and Gemini for business on web, desktop and mobile.
That is the news. It is not the decision.
The number that matters has a date on it
Argon’s launch price is $2 per million input tokens and $10 per million output tokens. Google’s pricing page carries that rate with an explicit qualifier: promotional “through November 30, 2026”. Immediately beneath it is the rate that follows, “starting December 1, 2026” — $4 input, $20 output. Batch processing follows the same doubling, from $1/$5 to $2/$10.
Vendors usually announce an introductory price and leave the reversion vague, which gives them room to quietly make the promotion permanent. Google did the opposite: it published both numbers and the switchover date at launch. When a vendor tells you the end date up front, the end date is the plan.
So the honest way to read this launch is that Gemini 4 Argon costs $4/$20, and for 46 days Google is paying half of it for you.
Who Google actually priced against
Line the two rates up against the models a buyer would really be choosing between:
| Model | Input /M | Output /M | Tier |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 | frontier |
| Claude Opus 5 | $5.00 | $25.00 | frontier |
| Gemini 4 Argon — from 1 Dec | $4.00 | $20.00 | frontier |
| Claude Opus 5.5 | $4.00 | $20.00 | frontier |
| Gemini 4 Argon — until 30 Nov | $2.00 | $10.00 | frontier, promotional |
| Claude Sonnet 5.5 | $2.00 | $10.00 | mid |
| GPT-6 Sol | $2.00 | $10.00 | mid |
Two exact collisions, and neither looks accidental.
Until 30 November, Argon is priced to the decimal at what OpenAI and Anthropic charge for their mid-tier models. A team evaluating GPT-6 Sol or Claude Sonnet 5.5 right now can swap in a frontier-class model at identical token cost. That is an aggressive, time-boxed bid for exactly the evaluations happening this quarter.
From 1 December, Argon is priced to the decimal at Claude Opus 5.5. Google has not undercut Anthropic at the frontier — it has matched it. The promotional window is the entire price advantage, and it expires.
The output ceiling is a new category of bill
The specification change that will surprise people is not the context window. Argon’s input context is 1M tokens, which is now unremarkable — Claude, the GPT-6 family and Gemini Flash all run 1M or more.
It is the output limit. Google raised Argon’s maximum decode to 1,048,576 tokens, up from a 64K ceiling, and the models page describes it as “long decodes up to 1M tokens”. Output is the expensive half of every rate card, and this ceiling moved exactly sixteen-fold. Note that it is opt-in: the model’s documented default output limit is still 65,536 tokens, so the long decode is a setting you turn on, not a behaviour you inherit.
Run that through the meter:
| Model | Max output | Cost of one maximum-length response |
|---|---|---|
| Gemini 3.8 Flash | 65,536 | $0.25 |
| Gemini 4 Argon (until 30 Nov) | 1,048,576 | $10.49 |
| Gemini 4 Argon (from 1 Dec) | 1,048,576 | $20.97 |
A single API call that can bill twenty dollars in output is new. For the long-horizon agent workloads Google is explicitly targeting — code migration, enterprise research, cyber defence — this is the parameter to cap in code before the first production run, not the one to discover in a monthly invoice. Set max_output_tokens deliberately. The default ceiling is no longer a safety net.
The caching asterisk
Google’s $2/$10 figure comes with cached input discounted 95%, which sounds like it rescues the cost model for repeated-prefix workloads. Read the mechanism: the pricing page states that context caching on Argon supports implicit caching only.
Implicit caching means Google applies the discount when it recognises a repeated prefix. You cannot pin a cache, reserve one, or guarantee a hit rate. That is a meaningfully weaker instrument than the explicit caching available elsewhere, and it means the 95% figure is not something a buyer can design around. Budget on uncached input; treat hits as upside.
Every number Google has published is Google’s
The capability case for Argon rests entirely on first-party evaluation. Google’s 30 September announcement reported 77.9% on DeepSWE v1.1, 51.3% on AutomationBench for a claimed first place, 91.7% on LVBench, and 68% on CWE-bench v1 for a tie at the top. Those are good numbers. They are also the vendor’s own, published alongside a launch and ahead of general availability.
Update, later on 15 October: the first independent read has landed, and it is less flattering. Artificial Analysis scores Gemini 4 Argon at 53 on its Intelligence Index, tested at high — which is Argon’s top thinking level — against 58 for Claude Opus 5.5, the model Argon will be price-matched against from 1 December, and 56 for mid-tier Claude Sonnet 5.5. Argon lands in a three-way tie with GPT-6 Astra and Claude Fable 5.1. The full cost-and-capability arithmetic is in Gemini 4 Argon vs Claude Opus 5.5.
That ranking matters more than usual here because of what the previous independent read said about Google. Artificial Analysis Intelligence Index v4.2 placed Google seventh among labs on frontier capability — a ranking taken while Google had nothing at the frontier to rank. Argon is the model that is supposed to change that position, and the evidence that it does is currently a slide from the company selling it.
There is a second reason to wait for third-party numbers: the model that was benchmarked in September was the Fairwind build, served to vetted cyber-defence customers. The model on sale from 14 October is the general-availability build. Vendors do not usually publish deltas between the two, and nothing in Google’s announcement says the scores were re-run against what a paying customer now calls.
None of this is a reason to skip Argon. It is a reason to run your own evals during the promotional window rather than to procure against a press release — which happens to be exactly what the next 46 days are good for.
What a buyer should do in the next 46 days
Argon is a real frontier model that a real customer can now buy, which is a genuine change in Google’s position and the first time in 2026 that the Gemini line has had an answer at the top of the curve. The capability claims put it mid-pack among frontier models rather than ahead of them, but the access question — the one that actually blocked procurement — is resolved.
The timing question is not.
- Evaluating this quarter: run the evaluation now, on the promotional rate, and price the production decision at $4/$20. A pilot that only works at $2/$10 is a pilot that fails on 1 December.
- On Claude Opus 5.5: from December the token prices are identical, so the comparison reverts to capability, output ceiling, caching mechanics and lifecycle terms — and on the last of those, Google publishes no minimum notice period while Anthropic commits to 60 days.
- On GPT-6 Sol or Claude Sonnet 5.5: you have a 46-day window to test whether a frontier model earns its place in your stack at no extra token cost. That is the most valuable thing about this launch and it has a hard deadline.
- Running agents: cap maximum output tokens before anything reaches production.
- Budgeting annually: use $4/$20. Every annual figure built on the launch price is half the real number.
Google spent 2026 shipping Flash models while its Pro line stalled and its frontier model sat behind a clearance programme. Argon on general availability closes that gap. The rate card says it closes it at Anthropic’s price, six weeks from now.
Frequently asked questions
What does Gemini 4 Argon actually cost?
Two different things depending on when you read your invoice. Through 30 November 2026 it is $2 per million input tokens and $10 per million output tokens, with batch processing at half that ($1/$5) and cached input discounted 95%. From 1 December 2026 the standard rate is $4 input and $20 output, with batch at $2/$10. Both figures are published on Google's own pricing page, with the switchover date stated in the same note — this is a scheduled doubling, not a risk that prices might rise.
Is the introductory price a discount or is the later price the real one?
Treat $4/$20 as the real price. Google published both numbers and the date simultaneously, which is the clearest possible signal that $2/$10 is a launch promotion rather than a rate it intends to defend. Any cost model, annual budget or build-versus-buy comparison built on $2/$10 is wrong by a factor of two from 1 December. The practical use of the next 46 days is evaluation, not deployment at scale — unless your workload can move vendors quickly if the arithmetic stops working.
How does Argon compare on price to Claude and GPT?
During the promotion, $2/$10 is identical to GPT-6 Sol and Claude Sonnet 5.5 — both mid-tier models. So for 46 days Google is selling a frontier model at mid-tier rates. From 1 December, $4/$20 is identical to Claude Opus 5.5, Anthropic's current frontier model. Google has not undercut Anthropic at the frontier; it has matched it, after a window in which it undercuts everyone. GPT-6 Astra remains more expensive at $10/$50.
Why does the 1M-token output limit matter for cost?
Because output is the expensive half and the ceiling just moved 16-fold. Argon can emit up to 1M tokens in one response, raised from 64K. At $10 per million output tokens a single maximum-length response bills $10.49; from 1 December it bills $20.97. The long decode is opt-in — Argon's documented default output limit is 65,536 tokens — so the exposure arrives only when you raise the cap. By comparison, Gemini 3.8 Flash caps output at 65,536, so its worst-case single response costs about $0.25. If you let an agent loop over long decodes without a max-output cap set in code, Argon is the first Google model where one runaway call is a material line item.
Can I rely on the 95% cached-input discount?
Only partly. Google's pricing page states that context caching on Argon supports implicit caching only, meaning the discount applies when Google recognises a repeated prefix rather than when you explicitly pin one. You cannot reserve a cache, guarantee a hit rate, or design a system around a committed cached rate the way explicit caching allows elsewhere. Model your costs on uncached input and treat cache hits as upside.
Sources
- Google Cloud Blog — Gemini 4 Argon available on Google Cloud (primary, 14 October 2026: general availability 'starting today'; model ID gemini-4-argon; introductory $2/M input and $10/M output through 30 November 2026, cached input 95% off; $4/M and $20/M thereafter; available in Gemini for business on web, desktop and mobile, the Gemini platform API, Google AI Studio and Google Antigravity)
- Gemini API Docs — Pricing (primary, checked 15 October 2026: Gemini 4 Argon at $2.00/M input and $10.00/M output promotional 'through November 30, 2026', $4.00/M and $20.00/M 'starting December 1, 2026'; batch at $1.00/$5.00 promotional and $2.00/$10.00 after; context caching supports implicit caching only)
- Gemini API Docs — Models (primary, checked 15 October 2026: gemini-4-argon described as 'our flagship reasoning model with a unified 1M context window and long decodes up to 1M tokens'; the only Gemini 4 family member listed)
- Google Blog — Gemini 4 Argon: our next era of frontier intelligence (primary, 30 September 2026 announcement: DeepSWE v1.1 77.9%, AutomationBench 51.3% ranked first, LVBench 91.7%, CWE-bench v1 68% tied for first; output token limit raised to 1M from 64K; initial access through the Fairwind programme for trusted cyber defenders)
- Claude Platform Docs — Pricing (primary, checked 15 October 2026, for comparison: Claude Opus 5.5 at $4/M input and $20/M output; Claude Sonnet 5.5 at $2/M and $10/M)
- Artificial Analysis — Models leaderboard (independent, checked 15 October 2026: Gemini 4 Argon 53 on the Intelligence Index at high; Claude Opus 5.5 58 at max with fallback and 56 at xhigh; Claude Sonnet 5.5 56; Claude Fable 5.1 53; GPT-6 Astra 53; GPT-6.1 Sol 52; Grok 4.7 46)
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.