PANONDA
Últimas
← News
Ferramentas·3 min read

Google's cheap Gemini model went public at $2/$10. Here's what that actually buys you

Gemini 4 Carbon is generally available through the Gemini API at $2 per million input tokens and $10 output. Google says it undercuts rivals by 4-5x. The comparison isn't verified — but the price is, and here's the math on a real publishing workflow.

By PANONDA Newsroom

Foto: Markus Winkler · Pexels

What happened

Google made Gemini 4 Carbon generally available on November 10, through the Gemini API in AI Studio and Gemini for business. The listed introductory price is $2 per million input tokens and $10 per million output tokens.

Google also says Carbon runs roughly one-fifth the cost of OpenAI's GPT-6 Astra and about one-quarter the cost of Anthropic's Fable 5.2. Those comparison figures come from Google. They were not independently confirmed, and neither were Carbon's benchmark numbers.

So the defensible version is narrow: a capable model is now publicly available at an unusually low posted price. That's it. Not "the best model." Not "cheapest forever."

The math on a real workflow

At the listed rates, here is what you'd actually pay.

  • 10 million input tokens: $20
  • 10 million output tokens: $100
  • 100 million input tokens: $200
  • 100 million output tokens: $1,000

Now put that on a show. A 60-minute podcast transcript runs roughly 12,000 tokens. Add your prompt, feed in about 20,000, get 4,000 back — show notes, chapters, three social cuts. That's about eight cents an episode.

Weekly show, full year: under five dollars.

A heavier job — 1 million input tokens and 200,000 output, say a month of articles converted into platform-specific headlines, descriptions and newsletter variants — comes to roughly $4 at the posted rates.

If you are a one- or two-person shop, the entire price difference between Carbon and the expensive models is smaller than your podcast hosting bill. The tokens were never what was stopping you.

What this is not

It's not a reason to migrate a working pipeline. Switching models costs you prompt tuning and a quality regression you find three weeks later, and you don't recover that in four-dollar savings.

It's not a confirmed permanent rate. The word is "introductory." We've already seen one Gemini tier listed at $2/$10 with a scheduled jump to $4/$20 on December 1, according to pricing trackers. Nothing in the available reporting says how long Carbon's price holds.

And it's not a complete price sheet. No confirmed caching discounts, rate limits, image or video pricing, regional pricing, enterprise minimums or service-level guarantees. Video and image inputs in particular burn tokens at a completely different rate than text — a pipeline that tags b-roll is not a five-dollar-a-year problem.

Who this actually moves

Three situations where cheap tokens change a decision.

Archive jobs. Running a 2,000-episode back catalog through a model to build a searchable index puts you in the hundreds of dollars at flagship pricing and the tens at this one. That's a real line item, and it's a one-time job you can schedule.

Translation at volume. First-pass translations for human editing scale linearly with word count. If you publish in three languages, the per-million rate is your whole budget.

Anything customer-facing. A chatbot, a search box, an AI feature in your app — your cost scales with audience, not with your output. A viral week is an expensive week. That's the only case where the posted rate should drive the decision.

What to do this week

  1. Pull last month's actual API bill. If it's under $50, you don't have a model-pricing problem and you can skip this entire news cycle.
  2. Run one job both ways. Same prompt, same input, Carbon versus whatever you use now. Compare output quality and latency before you compare invoices.
  3. If you have a bulk one-time job scoped — archive transcription, catalog tagging, full-library summarization — queue it while the introductory rate holds.
  4. Check Google's own pricing page before you budget. The $2/$10 figure is in the reporting; the fine print on caching, limits and expiry is not.
  5. Set a hard spend cap at the provider level on anything customer-facing. Usage billing without a ceiling is how a good traffic day becomes a bad invoice.
  6. Treat the "one-fifth of GPT-6 Astra" line as a vendor claim, because that's what it is. Price your own job three ways.

Sources

Everything above was checked against these pages. Open them and see for yourself — that is why they are here.

  1. 1theverge.com
  2. 2csuite.so
  3. 3modelpricewatch.com
Share

Get the stories by email

What changes in AI for people who publish — sourced, and to the point.

No spam. Unsubscribe anytime.