PANONDA
Últimas
← News
Ferramentas·3 min read

Gemini 4 Carbon's $2/$10 price ends November 30. Here's the math before and after

Google's cheap model went public at $2 per million input tokens and $10 output. Reporting now puts an expiry on that: December pricing doubles to $4/$20. That changes the order you run your jobs in, not which model you use.

By PANONDA Newsroom

Foto: Tim Heckmann · Pexels

What happened

Google made Gemini 4 Carbon publicly available on November 10 at $2 per million input tokens and $10 per million output tokens. When we covered the launch, no reporting said how long that rate would hold.

Now there's a date. According to VentureBeat, the launch rate runs through November 30, after which Carbon moves to $4 per million input tokens and $20 per million output tokens. That figure comes from VentureBeat's account; a primary Google pricing page was not available in the reporting we could check.

For comparison, Anthropic lists Claude Fable 5.2 at $8 input / $40 output — itself a 20% cut from the previous version, per Fortune.

The math, two ways

Take a workload of 10 million input tokens and 2 million output tokens. That's roughly a month of a mid-size archive job: feed in articles or transcripts, get back summaries, tags and headline variants.

  • Carbon, launch price: about $40
  • Carbon, December price: about $80
  • Fable 5.2, list: about $160

So the deadline is worth $40 on that job. The gap to Anthropic's list price is worth $80 either way.

The expiry doesn't change which model is cheapest — it changes whether a one-time bulk job is worth scheduling this month instead of next. That's the whole story.

What this is not

It is not a 75% discount on your AI bill. Those are token charges only. They exclude storage, retrieval, embeddings, orchestration, tool calls, human review and any vendor minimums. If your pipeline has a person reading the output — and it should — the person is the budget.

It is not a quality claim. No benchmark in the available reporting shows whether Carbon matches Fable 5.2 on long-form editing, factuality, multilingual work or code. Carbon's context window, rate limits and batch pricing also weren't confirmed. A cheap model that needs two passes isn't cheap.

It is not a reason to migrate. Anthropic's batch tier is reported at $4/$20 — the same as Carbon's December rate. If your jobs are asynchronous and you already run on Claude, the price gap may be smaller than the list prices suggest.

And it is not urgent for most of you. If your monthly API bill is under $50, the November 30 date costs you a cup of coffee. Skip it.

Who the deadline actually moves

One-time bulk jobs. Back-catalog transcription, library tagging, a full-archive summarization pass. These are scheduled, not continuous. If the job is already scoped, running it before November 30 is free money.

Translation at volume. Cost scales linearly with word count. Three languages across a full publication is where the per-million rate becomes the budget.

Customer-facing features. Chatbots and AI search boxes scale with audience, not output. A good traffic week is an expensive week. Here the posted rate matters every month, not just this one.

What to do this week

  1. Pull last month's actual API invoice. Under $50 and this news cycle doesn't apply to you.
  2. List any one-time bulk job you've already scoped — archive, catalog, translation backlog. Those are the only things worth rushing.
  3. Run the job on a 5% sample first. Check output quality before you commit the full volume at any price.
  4. Queue the full run before November 30 if the sample passes.
  5. Rebuild your December budget at $4/$20, not $2/$10. Assume the increase happens.
  6. If you're on Claude, price your job against batch rates, not list rates, before you conclude you're overpaying.
  7. Check Google's own pricing page before you sign anything. The expiry date is in the reporting, not in a document we could verify.
  8. Set a hard spend cap at the provider level on anything customer-facing. Usage billing without a ceiling is how a viral day becomes a bad invoice.

Sources

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

  1. 1venturebeat.com
  2. 2modelpricewatch.com
  3. 3fortune.com
  4. 4local-ai-zone.github.io
Share

Get the stories by email

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

No spam. Unsubscribe anytime.