Gemini 4 Carbon, the checkpoint Google appears to be lining up behind Gemini 4 Argon, has started reaching people outside the company, barely a week after Argon went public. TestingCatalog collected the first impressions on 23 October 2026. They come from early-access testers and arena matchups, so they are not an official release and not benchmarks.
An X user going by Lentils wrote on 21 October that a Carbon early access program had begun, with partners getting the model and LM Arena (the public site where visitors vote between anonymous model answers) running it. Lentils expects a launch very soon and says Carbon is clearly stronger than Argon at 3D work. Frontend and interface output is still bad, and the model is slower, though Lentils calls it “less lazy.”
A second tester, Lumina, posted most of the sample outputs. Lumina says Carbon showed up on Arena on 21 October and sees “noticeably higher quality than Argon does” in a voxel pagoda comparison against other frontier models. The price is speed: Lumina calls the model “incredibly slow,” about two to four times the wait of Argon, varying by task. A pelican-on-a-bicycle SVG test showed the same weak interface work and the same lag, yet Lumina says Carbon “feels good” on the many other tasks tried.
The leak is awkward for Google. A leaker named Lyra posted a screenshot of the documentation for the Gemini API Early Access Program, a closed circle of developers, researchers, and partners who see upcoming models ahead of launch. According to TestingCatalog, the Trusted Tester Agreement tells participants to keep unreleased models, their codenames, and any results private until launch, to keep the models out of production, and to send evaluations to Google directly. Carbon samples circulating openly on X cut against every one of those terms.
Treat the numbers with care. TestingCatalog points out that these are head-to-head arena votes on prompts the testers chose themselves, not benchmarks, and that the checkpoint behind any given answer is unknown. The two-to-four-times slowdown comes from one tester’s impression, and no one has published a latency measurement.
The pace matters to Google as much as the model does. Logan Kilpatrick posted that Argon pushed about 25 trillion tokens through Google’s systems within its first three days, and TestingCatalog suggests the company will want a quick follow-up before the hype fades. The same outlet recently noticed wider Computer Use permissions inside the Gemini desktop app. It also says Google seems to be basing Nano Banana 3 Pro on Gemini 4, aiming at a November to December window.
Rivals are moving too. TestingCatalog says GPT-6.1 Terra, which OpenAI is readying, turned up in Microsoft code. Anthropic’s Fable 5.5 has begun going out to partners, and further Anthropic models look likely. A Carbon that fixes Argon’s 3D weakness without a long wait would give Google a strong reply, which is the condition TestingCatalog itself sets.
Lentils and Lumina agree on the trade: better 3D, worse speed, same weak interface output. A model that answers two to four times slower fits batch jobs and offline generation long before it fits a live chat window, so anyone planning around Carbon should wait for real latency numbers before the quality screenshots decide anything.
Reported by TestingCatalog on 23 October 2026.
