Tool of the Week: Gemini 4 Argon, and the price that has an end date
Google's Gemini 4 Argon reached the Gemini API this week, and the price is the story. Launch pricing is $2 per million input tokens and $10 per million output. Per a footnote in Google's launch post, as quoted by an analysis site, the price after the introductory period is $4 and $20. Coverage puts the end of the promotion at November 30. That is double, on both sides.
If you are only trying the model, ignore this. If you are building something on it, this is a budget line. A pipeline you size at the launch price will cost twice as much in December without a single code change.
A rough example, with round numbers I made up: a job that sends 1 million input tokens and gets 200,000 tokens back costs about $4 at the launch price ($2 for input, $2 for output) and about $8 at the standard price. Run that daily and the gap is real money by the end of a quarter.
One billing detail to check before you estimate: the same analysis says the output price includes thinking tokens, so the model's reasoning is billed at the output rate. A short answer can still be a long bill. Google had not published long-context, batch or cache-storage pricing for Argon in the coverage I read, so those are unknowns, not zeros.
Context is about one million tokens combined input and output. Google's benchmark claims are Google's own and, per the coverage, mixed against competitors, so test on your own data.
What I would do: build the cost model at the standard price, not the launch price, and treat anything under it as a bonus. If the numbers only work at $2 and $10, you have a promotion, not a plan.
Quick Hits
Microsoft released its own decision model, a follow-up to last week's story. Microsoft-Decision-1 was announced October 9. It is a small text-only model built on Alibaba's Qwen3.5-9B that returns a probability for each fixed answer, aimed at routing, classification and checking agent actions. It is available in Microsoft Foundry and through OpenRouter at $0.042 per million input tokens, with output free. The speed and accuracy figures are Microsoft's own. Microsoft also says it should not be the sole basis for consequential decisions about people, such as credit or hiring.
A voice model alias changed under people's feet. Release trackers list xAI's Grok Voice 3.0 as out October 14, with the API unchanged and the grok-voice-latest alias now pointing at it. That is a quiet way to change production behavior. If you call any model through a "latest" alias, pin a specific version, or accept that your output can change on someone else's schedule. I am relying on trackers here, not an xAI post.
Prompt of the Week: The Cost Both Ways
Help me estimate what a recurring AI workload will cost at a
launch price and at the standard price that follows.
Workload: [what it does, e.g. summarize 500 support tickets daily]
Runs: [how many times per day or week]
Average input per run: [tokens, or pages/words if unsure]
Average output per run: [tokens, or words if unsure]
Reasoning or thinking is billed as output: [yes / no / unknown]
Launch price per million tokens: input [x], output [y]
Standard price per million tokens: input [x], output [y]
Do this in order:
1. Convert my numbers to tokens, and say which conversions are
guesses. Use roughly 4 characters or 0.75 words per token.
2. Calculate the cost per run, per month and per quarter at each
price. Show the arithmetic.
3. Add a second scenario where output tokens are 3 times higher
because of reasoning.
4. Tell me which numbers I should replace with measured values
from my own logs before I trust this.
5. State the monthly cost at which this workload stops being
worth automating, based on what a person would cost for the
same work. Ask me for that figure if I did not give it.Fill it in with a real workload, then replace the guesses with numbers from your own usage logs. The estimate is only as good as the token counts.
One last thing
Writing this is one side of the work. The other is building it: data pipelines, file and reporting workflows, integrations between systems that were never meant to talk, and the automation that takes the manual steps out.
If a process in your week still depends on someone remembering to run it, book a free 15-minute audit and I will find your top 3 time-wasters. If it is not worth automating, I will tell you that.
Scott