Carbon is the cheapest frontier model on paper. No one has benchmarked it on publishing work
Google's Gemini 4 Carbon lists at $2 per million input tokens and $10 output until November 30. The discount is real. The proof that it does your job as well as what you run today does not exist yet — so use the window to test, not to migrate.
By PANONDA Newsroom

Foto: Lukas Blazek · Pexels
What happened
Google put Gemini 4 Carbon into public release on November 10 at $2 per million input tokens and $10 per million output tokens. It's reachable through the Gemini API in AI Studio and Google's business products. Google AI Ultra subscribers get it inside the Gemini app, AI Mode in Search, and Antigravity.
Google says the introductory rate is roughly one-fifth the cost of OpenAI's GPT-6 Astra and one-quarter the cost of Anthropic's Fable 5.2. That's the company's own comparison, not an independent one.
The rate has an end date. Reporting puts the promotion's close at November 30, with Carbon moving to $4 per million input tokens and $20 per million output tokens after that. We could not verify every billing condition against a primary Google pricing page.
What the discount does not tell you
There is no public benchmark showing Carbon is better than GPT-6 Astra or Fable 5.2 on the work publishers actually do: long-form editing, factual accuracy over a transcript, multilingual rewriting, metadata that holds up.
A per-token price is not a per-job price. Two models billed at different rates can consume wildly different numbers of tokens on the same task. If Carbon needs more context to get there, or returns longer answers, or needs a second pass because the first one was wrong, the cheaper listed rate evaporates. The number that matters is cost per published asset, and only your own pipeline produces it.
Carbon's context window, rate limits and output caps weren't confirmed in the reporting we checked. Don't plan a long-document workflow around assumptions.
The math, so you know the stakes
Token billing only — no storage, retrieval, orchestration, tool calls or human review.
- 1M in + 1M out: $12 now, $24 in December.
- 10M in + 2M out: $40 now, $80 in December.
- 100M in + 10M out: $300 now, $600 in December.
That last line is a real newsroom-scale job: a full archive pass for summaries and tags. The deadline is worth $300 on it. On the first line, the deadline is worth twelve dollars. Most of you are closer to the first line than the third.
Where the savings actually show up
Not in editorial judgment. In the repeatable, high-volume work where output quality is easy to measure: tagging, transcript cleanup, translation, archive search, headline variants, versioning for different platforms. You can score those against a rubric. You cannot score "is this a good column" against a rubric, which is why the first credible savings are operational, not creative.
The discount window is a free testing budget. That's its best use. Cheap inference is the condition that lets you run the same 200 articles through two models and compare, which nobody does at full price.
What to do this week
- Pull last month's API invoice. Under $50 and none of this applies — skip to step 8.
- Pick one repeatable job you already run: transcript cleanup, tagging, or translation. One, not three.
- Build a 50-item test set from real work you've already published, with the correct output you'd accept.
- Run it on Carbon and on your current provider. Log four things: accuracy, latency, tokens consumed, and minutes of human editing after.
- Compare on cost per finished asset, not on the posted rate. Editing time is the line item that decides it.
- If a one-time bulk job already passed a sample test, queue it before November 30. Nothing else needs rushing.
- Rebuild your December budget at $4/$20. Assume the increase happens and be pleasantly surprised if it doesn't.
- Set a hard spend cap at the provider level on anything customer-facing. Usage billing with no ceiling is how a good traffic 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.
Get the stories by email
What changes in AI for people who publish — sourced, and to the point.
No spam. Unsubscribe anytime.