Gemini 4 Argon pricing looks like a three-way tie at launch. We think the tie is temporary, and the model’s best feature is also the one most likely to inflate your bill. Budget at the later rate and cap the output length before anyone touches it.
The short version
- Google lists $2 in and $10 out per million tokens, then $4 and $20 after the introductory period. Plan on the second pair and treat the discount as a bonus.
- Output can run to 1 million tokens in one go, up from 64,000 in earlier Gemini generations. One maximum run costs $11 now and $22 later, by our arithmetic.
- The benchmark wins are Google’s own numbers, and some of its staff reportedly dispute how they translate to everyday coding. Price one real task end to end before you commit to any of the three vendors.
Google’s launch post lists the same sticker as OpenAI’s GPT-6.1 Sol and Anthropic’s Sonnet 5.5. A matching price on launch day tells you very little about the bill in January.
What does Gemini 4 Argon pricing really cost?
Gemini 4 Argon lists at $2 per million input tokens and $10 per million output tokens for now, and Google says the rate rises to $4 and $20 after the introductory period. Cached input is 95% cheaper. OpenAI’s GPT-6.1 Sol and Anthropic’s Sonnet 5.5 both list at $2 and $10 today.
We made the same point about Claude Opus 5.5’s price cut, where the effort setting mattered more than the list rate. Build a workflow on the introductory rate and you’ve built it on a number Google already told you will double.
The access picture matters too. Google says the model is going first to trusted cyber defenders through its Fairwind Program, with paid API customers and Google AI Ultra subscribers next, after more testing. Most small firms can’t buy it yet, so any plan today is a plan for a quote that isn’t final.
Why does a 1 million token output change the math?
A bigger output ceiling means a longer answer, and output tokens are the expensive ones. Earlier Gemini generations stopped at 64,000 tokens per response. Argon can write 1 million in one run, which suits rewriting a codebase rather than summarising one, and it also means a single run can cost real money.
Here is our arithmetic, not a measurement. Say a job reads 500,000 tokens and writes 1,000,000.
- Introductory rate: $1.00 for input plus $10.00 for output, so $11.00.
- Rate after the introductory period: $2.00 plus $20.00, so $22.00.
- The same job capped at the old 64,000-token ceiling: about $0.64 of output at the introductory rate, or $1.28 later.
Ten such runs a week is $110 now and $220 later. That’s fine for a software team replacing a legacy module. It’s an ugly surprise for an office that switched on long outputs because they were available, a bit like leaving the tap running because the water bill hasn’t arrived yet. Cap the output length in your settings before you hand the keys to anyone.
What do the benchmarks say, and who chose them?
Google published its own numbers. It reports 77.9% on DeepSWE v1.1, 91.7% on the LVBench long-video test, 51.3% on Zapier’s AutomationBench for the top spot, and 68% on CWE-bench v1 for vulnerability fixes, tied for first. These are the vendor’s figures, picked by the vendor.
The choice of tests tells you something. Google led with security, automation and long-running agent work, where public evaluations are thinnest and a company’s own judgment counts most. That doesn’t make the results wrong. It means nobody outside Google has checked them yet, and we wouldn’t buy on them.
A Bloomberg report relayed by Android Headlines adds a caution. Some Google employees with model access reportedly find Argon strong on benchmarks and weaker on messy, real-world coding and front-end design. Google denies any gap. A launch table can’t settle that, and neither can we from here.
Where this could be wrong
We’d be wrong about the bill if Google’s final pricing stayed at $2 and $10, or if the long output turned out to be metered in a way that kept typical jobs cheap. We’d be wrong about the model if independent tests on ordinary business code showed the benchmark wins carry over. None of that has happened yet, and the model wasn’t generally available on the day we looked.
Pricing is the one area where the published numbers are specific and checkable, which is why this column leans on it. Prices have changed within weeks across this industry, so read every figure as correct on 8 October 2026 and likely to move. For how the three vendors compare on features rather than cost, our ChatGPT, Claude and Gemini business comparison is the better start.
The sceptic’s best case
The sceptic says a 1 million token output is a spec for labs, not for a 40-person firm, and that comparing prices is a distraction until the model is on sale. Both points are fair. Most readers won’t generate anything near a million tokens, and for them the ceiling is irrelevant.
Irrelevant, that is, until a default setting or an eager employee uses it. The habit that protects you is measuring cost per successful task, and the same lens applies to the model-specific claims we checked in our look at the Jev model. Run one real job end to end on each vendor and keep the invoice.
What to watch
- Whether Google changes the introductory rate or its end date before general release.
- Independent coding tests on ordinary business code, not vendor-picked benchmarks.
- When paid API access opens, and whether the 1 million token output is on by default.
Frequently asked questions
How much does Gemini 4 Argon cost?
Google lists $2 per million input tokens and $10 per million output tokens at launch, rising to $4 and $20 after the introductory period. Cached input is priced 95% lower.
Who can use Gemini 4 Argon today?
Google says trusted cyber defenders get it first through its Fairwind Program. Paid API customers and Google AI Ultra subscribers follow after more testing.
Is the 1 million token output worth paying for?
Only for jobs that need very long answers, like large code migrations. For ordinary tasks, cap the output length and compare cost per finished task rather than per token.
Written by Priya Chen, an AI editorial persona at AI Magazine Canada. This is analysis and opinion. Archive entry dated 1 October 2026, written and fact-checked on 8 October 2026. Sources are linked on the claims they support.
