Gemini 4 Carbon tops an independent AI ranking, but almost nobody can use it yet

A top score with no way in

On 28 October 2026, Google DeepMind published a post introducing Gemini 4 Carbon, which it calls its next frontier model. The same day, the independent evaluator Artificial Analysis added Carbon to its Intelligence Index v5.0 with a score of 59. That puts it first, one point above Claude Opus 5.5 at 58 and nine above Gemini 4 Argon, the model Carbon is built on.

As OrcaRouter reported, though, the gap between the ranking and real access is the actual story. Argon has been sold on Google's paid API since 13 October. Carbon, as of 30 October, could not be called by anyone outside a named programme.

This is not a leak or a rumour. The model is announced, benchmarked and priced. What is missing is the practical part: a way for ordinary developers to use it.

What Google shared and what it held back

Google describes Carbon as a model for cybersecurity and long, multi-step engineering work. Koray Kavukcuoglu, the company's chief AI architect, is named in the post. Access begins with a subset of cyber defenders and then widens to trusted partners through something called the Fairwind Program. Developers, businesses and consumers come after that, once safeguards are refined. Google also says it is taking part in the U.S. government's voluntary pre-release testing process. No dates are given for any step.

Several details a launch normally includes are missing:

• A model ID, the exact name developers type to call a model. Google's API list on 30 October showed only gemini-4-argon.
• A context window, meaning how much text the model can consider at once. Artificial Analysis lists 1M, but that is its figure, not Google's.
• Size and design details, plus any technical report.
• A model card, the standard document describing a model's abilities and limits.

Pricing is published in full, which makes the gaps look like a choice rather than an oversight.

Reading the 59 more closely

The index averages nine separate tests. Carbon wins four of them: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. So the top spot comes from being strong nearly everywhere, not from a clean sweep.

On coding agents, which are AI tools that write and run code on their own, Carbon scores 59 using Google's Antigravity CLI. Claude Opus 5.5 reaches 62 inside Claude Code. Opus 5.5 also beats Carbon on Terminal-Bench 4.0, where Carbon has 64%, and on FrontierCode v1.1, where Carbon has 51%.

One result Google's post does not mention concerns accuracy. On AA-Omniscience, Carbon's accuracy rose 10 points to 58%. But its hallucination rate, how often it states something false with confidence, rose from 4% to 11%.

Google also published its own results. These include 88.2% on SEC-bench Pro, a test of finding software vulnerabilities, and 54.8% on ExploitGym against 42.0% for Argon. Treat these as Google's claims, since nobody outside the company has reproduced them. Google's table also admits losses: FrontierSWE 2 went to GPT-6 Astra, and an offline part of OSWorld-2.1 went to Claude Opus 5.5.

A price with no end date

Carbon's launch pricing is $2 to send in a million tokens (small chunks of text that models read and write) and $10 to get a million back. Cached input, text the model has already seen, costs 95% less. Afterwards the rates rise to $4 and $20, but Google gives no date for the change.

Argon works differently. Its introductory rate has a firm end: 30 November 2026, with standard pricing from 1 December. Carbon's clock cannot start until customers can actually buy it.

Artificial Analysis also estimates the cost of finishing a task on its index. At standard prices it puts Carbon at $7.76, about 23% below the $10.04 for Claude Opus 5.5 at its maximum setting. GPT-6 Astra at maximum comes in lower, at $6.25. These are per-task figures, which matter because a model that costs less per token but reasons for longer can still cost more per answer.

What happens next

Argon offers the only precedent. Google announced it on 30 September 2026 and opened the paid API on 13 October, thirteen days later. Applying that gap to Carbon points to roughly 10 November. That is simple arithmetic from a single case, not a promise, and the government testing has to finish first.

There is another possibility. On 21 October, a third-party tracker of public model comparisons noticed a change in how the gemini-4-argon name behaved. Some developers wondered if Carbon's improvements were arriving early. This is unconfirmed. But if Google folds Carbon's training into the Argon name, the sign to watch would be a version update to a model you can already call, not a new name.

Meanwhile, the Gemini models you can use today sit a step lower. OrcaRouter lists Gemini 3.1 Pro Preview at $2 input and $12 output with a 1M context, and Flash models from $0.75 input.