Google calls Gemini 4 Carbon "our strongest model", and it tops the Artificial Analysis Intelligence Index at 59. As of Nov 1, almost nobody can use it.
TL;DR
- Carbon is #1 on an independent ranking and posts a headline cybersecurity score.
- Carbon is a showroom model: the price is listed in public, but access is invite-only, first defenders, then partners, and everyone else gets it "as soon as possible" with no date attached.
- While you wait, read the benchmark tables and price your prompts at both rates.
Ranked first, pitched at cybersecurity
Google announced the model on Oct 28 with a cybersecurity focus. Its headline number is "88.2% on SEC-bench Pro, autonomously uncovering complex flaws buried in real-world software".
Defenders, partners, then a promise
Access starts with "a subset of cyber defenders", then "trusted partners". Google describes "a deliberate approach to rollout", which is a reasonable thing to say about a model that finds software flaws on its own. Ars Technica's headline puts the same fact more bluntly: "...its latest frontier AI model that you can't use".
The price, on the other hand, is already on the Google announcement. The tag reads "$2 per million input tokens and $10 per million output tokens". Ars reports those are introductory rates that will double later.
That makes Carbon a showroom model: spotlight, velvet rope, price tag taped to the pedestal. Any launch that tops a ranking, publishes its rates and names its approved users fits the same description.
Read, price, and check one page
Read the benchmark tables so you know what you are waiting for, and price your prompts now so the bill holds no surprises on release day. Then keep an eye on Google's announcement page. Access has opened when it stops naming "a subset of cyber defenders" and "trusted partners" as the audience.
Try this: take a prompt with 1M input tokens and 200K output tokens. It costs $4 at the intro rate, $8 if both rates double as Ars reports.
The price is public. The release date is not.
Sources: Google blog (Oct 28, 2026), Ars Technica (Oct 28, 2026), Reuters (Oct 28, 2026). Verified Nov 1, 2026: no release date announced.
Dan writes the AI satire on super-intelligence.lol, where this piece first appeared.
Top comments (0)