Gemini 4 Carbon Wins Benchmarks, But Buyers Can’t Use It Yet

By Adam Pease

Gemini 4 Carbon Wins Benchmarks, But Buyers Can’t Use It Yet

Google DeepMind has released Gemini 4 Carbon, two weeks after it launched Gemini 4 Argon. Google says Carbon takes the top spot on the Artificial Analysis Intelligence Index v5, with large gains in coding, agentic work, and cybersecurity. Most enterprises cannot use it, though. Carbon is going to selected cybersecurity defense companies first, and Google has given no date for general availability.

This blog overviews the Google Gemini 4 Carbon announcement and offers our analysis.

Why Did Google Announce Gemini 4 Carbon?

Google says Carbon shows “the pace at which we’re pushing the frontier forward.” Carbon scores 59 on the Artificial Analysis Intelligence Index v5. That is up from Argon’s 50 and, according to Engadget, “just ahead of” Claude Opus 5.5 and GPT-6 Astra. Google says it built the model for “much harder, long-horizon engineering problems,” and says internal teams ran it for days at a time on multi-step engineering tasks.

Cybersecurity is the center of the launch. Google says it ran some security benchmarks without safety guardrails. On those tests, Carbon scored 88.2% on SEC-bench Pro, 71.0% on CWE-Bench, and 54.8% on ExploitGym, ahead of both GPT-6 Astra and Argon, as reported by The Decoder. Google also says Carbon cut the attack success rate on a Gray Swan prompt-injection test from 0.7% to 0.0%. These are Google’s own results. Some of them, including the Chrome team’s patch-writing tests, cannot be independently verified, according to The Decoder.

Access is staged. Selected cybersecurity defense companies get the model first. More trusted partners follow through Google’s Fairwind Program. Developers, businesses, and consumers come after that, once Google has gathered feedback and adjusted guardrails. Carbon is also going through the U.S. government’s voluntary pre-release testing. Argon took the same path and is now in the Gemini app and AI Mode in Search.

Carbon launches at Argon’s introductory price of $2 per million input tokens and $10 per million output tokens. Artificial Analysis cites a list price of $4 and $20, with the 50% introductory discount running through the end of 2026. Sources differ on how long the introductory price lasts, because Ars Technica reports that Google declined to give an end date for Carbon’s.

Analysis

At Aragon Research, we view Carbon’s leaderboard win as saying little about what enterprises can buy today. Only a small group of cyber defenders has access, and Google has not said when anyone else will. A model that cannot be deployed is not a procurement option, however high it ranks. Ars Technica summed up the launch with the phrase “you can’t use.”

The headline score also hides trade-offs. Carbon does not lead everywhere. On the Artificial Analysis Coding Agent Index it scores 59, against 62 for Claude Opus 5.5. On AA-Briefcase, the agentic knowledge work test, it trails Kimi K3.1 Pro, Muse Wonder, and Grok 4.7. Accuracy rose from 48% to 58% compared with Argon, but the hallucination rate rose from 4% to 11%. The Decoder notes that 11% is still the second-lowest rate in that ranking. Even so, the rate nearly tripled, and buyers in regulated workflows should weigh that more heavily than a nine-point index gain.

Cost is a better point of comparison than rank. Artificial Analysis puts Carbon’s cost per task on its index at $7.76, against $7.68 for Argon, $10.04 for Claude Opus 5.5 (max), and $6.25 for GPT-6 Astra (max). Those figures use list prices, and a model’s real cost per task depends on the task. Most of the numbers in this launch come from Google or Artificial Analysis and have not been reproduced by others.

For competitors, Carbon shows how close the frontier is. Ars Technica points out that every frontier model scores in the same range and that the gap between Carbon and the field is only a few points. When the top models sit that close together, a launch-day ranking is a poor way to choose a vendor.

What Should Enterprises Do About This News?

Enterprises should compare model vendors on what they can deploy and what a task costs, not on launch-day rankings. In practice, that means:

• Leave Carbon out of vendor selection until it is available to your organization.
• Ask Google for the access timeline, the Fairwind Program terms, and the end date of the introductory price.
• Compare cost per task and hallucination rate, not only index rank.
• Test any model on your own tasks before moving workloads to it.

Security teams that may qualify for early access should ask Google whether they are eligible. They should also review how a model with exploit-writing capability will be governed inside their environment.

Bottom Line

Gemini 4 Carbon tops a major index, but it is not a product most enterprises can deploy, and its headline score hides a higher hallucination rate and a gap behind Claude Opus 5.5 on coding agents. Enterprises should judge model vendors on availability, cost per task, and performance on their own work, and leave launch-day leaderboards out of buying decisions.