Gemini 4 Carbon tops many tests, but Google wants more safety work before developers, businesses, or the rest of us can try it.
A new model behind a locked door
Google announced a new AI model on October 28. It is called Gemini 4 Carbon, and Google calls it the strongest model it has. You cannot use it. For now, only a small group of companies that protect computer systems from attacks can get it.
That is an odd way to launch anything. Most of us run our lives on software, and this model seems very good at finding the weak spots in software. So why would a company keep its best work away from nearly everyone?
Good at finding holes, and at using them
Google points to three security tests. In the first, a model must find real bugs in real software on its own. Google says Carbon scored 88.2%. In the second, a model must write working fixes for security flaws, and Google reports 71.0%. Google puts Carbon first on both.
The third test explains the locked door. It checks whether a model can turn a flaw it found into a working attack within ten hours. Carbon scored 54.8%, again first by Google's count. The way I see it, a tool that finds the broken window for the homeowner will find it for the burglar too. Google says this strength calls for a "deliberative approach" to the release, which in plain words means slowly.
One detail caught my eye. Google ran several of these tests with the model's safety limits switched off. I would like to see the same tests with the limits on. That leaves the practical question: who gets this thing, and when?
“I like that Google is going slowly, but I will trust the safety claims once someone outside Google has tested the version people actually get.”
Who gets in, and what keeps it safe
The first users, starting the week of the announcement, are selected security experts. Google says Carbon comes with layered protection, including live defenses against misuse and monitoring of its step-by-step reasoning. Once the U.S. government's voluntary testing is done, Google plans to bring in more trusted partners through a program called Fairwind. Developers, businesses, and ordinary customers come last. Google says they will get it as soon as possible, and promises no date.
Ars Technica, a tech news site, guesses the wait will be a couple of weeks. That guess assumes Carbon is released like Gemini 4 Argon, the model before it. Argon itself came out just two weeks before Carbon. How much better can a model be than something that new?
Slim margins and one surprise
Outside testers mostly agree it is better, though not at everything. Artificial Analysis, a company that tests AI models, ranks Carbon first overall, with 59 points to Argon's 50. It says Carbon does not lead in coding or in office work. In coding it comes second, behind Claude Opus 5.5.
Google presents Carbon as a totally new model. Ars Technica calls it essentially a slightly better Argon. Carbon does beat Argon in every test Google gave it, but by slim margins.
The surprise came from a test of plain facts. Compared with Argon, Carbon's correct answers rose from 48% to 58%. Its false answers, stated as if they were true, rose as well, from 4% to 11%. I read that as Carbon answering questions that Argon would often skip. More answers bring more right ones and more wrong ones. For false answers, lower is better. Carbon's 11% is still the second-lowest rate of any model, behind only Argon, while two big rivals sit at 50%. Google's own testing seems to point the other way.
Results nobody outside can check
Google says Carbon makes fewer unsupported claims than Argon, meaning statements with nothing to back them up. That is a different test from the false-answer test, so a model can do better on one and worse on the other. Still, nobody outside Google can verify results like that, and Google cites several. Its Chrome team says Carbon did more than 10 times better at writing high-quality fixes than an unnamed rival model. It is hard to lose a race against an unnamed runner.
None of this makes Carbon a bad model. The Decoder, another tech news site, says outside tests back up Google's numbers. What stays with me is who waits in line. The people who fix the broken windows get the new tool first. How much that head start is worth depends on how long they stay the only ones with it.
Nov 2, 2026 · Sketchbook, p. 04
