Google Announced Gemini 4 Carbon at #1 on the New Index. There Is No Public API Yet, and Argon's Price Goes Up on December 1. Plan Margins on the Model You Can Actually Call.
Google announced Gemini 4 Carbon on October 28 and it landed at the top of the revamped Artificial Analysis Intelligence Index with a score of 59. Claude Opus 5.5 sits at 58, GPT-6 Astra at 57 and Gemini 4 Argon, Google's own previous flagship, at 50. The catch for anyone who ships products: Artificial Analysis notes Carbon has no publicly available API yet, and Argon, the model you can actually call, has a price increase scheduled for December 1.
What you can and cannot use
According to VentureBeat, Google is initially limiting Carbon to selected cybersecurity defenders in its Fairwind Program. Broad release is promised "as soon as possible," with no date. Antigravity head Varun Mohan said access will widen through Fairwind before it reaches Antigravity more broadly, also without a date. Argon is still the model customers can use today.
Google did announce pricing: $2 per million input tokens and $10 per million output tokens at an introductory rate, rising to $4 and $20 afterward, with a 95% discount on cached input. The end of the introductory period was not specified. Notice what that means for planning. A price on a model you cannot call is a statement of intent, and the intro rate comes with an expiry nobody has dated.
Read the scoreboard carefully
The headline number is one point of lead over Opus 5.5 on a benchmark that was just rebuilt. Google says Carbon posts the top score in 15 of 17 rows in its own comparison, which is a fair thing for a vendor to say about its own table. The independent numbers are more mixed:
- Terminal-Bench 4.0: 65.2% for Carbon against 57.1% for Argon.
- FrontierSWE 2: Carbon scores 61.2%, behind GPT-6 Astra at 65.5% and Opus at 62.3%.
- OSWorld-2.1 (offline subset): Carbon 81.5%, behind Opus at 87.2%.
- Coding Agent Index: Carbon 59 against Argon 53, while Claude Code with Opus 5.5 leads at 62.
So the model that tops the general index trails on the tasks a solo builder pays tokens for: fixing real code and driving a computer. Carbon's strengths are security tasks. ExploitGym is 54.8% against Argon's 42.0%, SEC-bench Pro is 88.2% and CWE-bench is 71.0%. That explains the first-access program.
There is also a number in the independent results that should make you slow down. On AA-Omniscience, Carbon has the highest score (54) but a hallucination rate of 11%, against 4% for Argon. A model that knows more and invents more is not automatically an upgrade for a customer-facing feature. I wrote on October 24 that you should stop picking a coding agent by leaderboard after one audit found 44 of 65 tasks in a popular benchmark could be passed by cheating. Same rule here: one index is a starting point, not a purchase order.
What this does to your margins
If you built on Argon, your plan does not change today, because nothing about Carbon is callable. What does change is that December 1 is five weeks away and Argon's higher rates take effect that day. I would stop reading Carbon news as a reason to wait and start reading it as a reason to price your fallback.
Concretely, take your current monthly token volume for Argon, split by input and output, and recompute at the new rates. Then run the same math for the cheapest model that passes your own test set. Your own test set, not a published one. Your usage pattern decides whether cached input matters, and a 95% discount on cached tokens only helps if your prompts really reuse long prefixes.
Where I could be wrong
Maybe waiting a few weeks is the right call. If Carbon opens broadly in November at the introductory $2 and $10, that is a lower price than the post-increase Argon rates, and the hallucination gap might not matter for your use case. Google also has a record of widening access. My objection is only to building a plan around a date nobody gave. If you have a feature where quality is the bottleneck and cost is not, putting your name on the Fairwind or Antigravity waitlist costs nothing.
What I'd actually do
This week I would build a tiny evaluation of 30 to 50 real requests from my product and score Argon, my current cheaper fallback and my current best-quality model on it. Then I would write down my December margin at Argon's new rate. When Carbon opens, I rerun the same set and compare, instead of deciding from a leaderboard. The score that matters is the one measured on my own traffic, and the one I can run in an evening.
Author
Lukas
@lukcombinator