Image: Google

Google has reportedly started early testing of a new model called Gemini 4 Carbon with a small group of outside partners. Leaker Lentils80, who has a mostly solid track record, said on X that the Gemini API Early Access Program for Carbon has officially begun and that the release is very close.

Gemini 4 Carbon EAP has officially started. Early access has been given to partners and it's being tested in LMArena. The release is very close.

It's noticeably better than Gemini 4 Argon at 3D, though frontend/UI is still bad. It's also slower but less lazy.

— Lentils (@Lentils80) October 21, 2026

Lentils80 also said Carbon is being tested on LMArena, a public site where people vote between two unnamed AI models. In that user's early impressions, Carbon is noticeably better than Gemini 4 Argon at 3D work. It still struggles with website and app design, and it is slower but less lazy. Google has not confirmed the name, the testing or a release date.

Lyra, a reliable leaker, shared what appears to be a screenshot of Google's own page describing the program. We could not find that page live on Google's developer site, so we cannot confirm where it came from.

Gemini 4 Cabrón pic.x.com/OLBxTYEngk https://t.co/Z7EdDF1hg2

— lyra (@lyraxana) October 21, 2026

The page explains how Google tests models before launch. It calls the program a collaboration with selected developers, researchers and partners, who get early access to upcoming Gemini API and Google AI Studio models. It says the program is separate from Google's Vertex AI early access programs. Entry is by invitation only, and Google picks testers for their technical skill, their use cases and how well they give feedback.

The rules are strict. Testers sign a Trusted Tester Agreement. It bars them from revealing that they are in the program, or sharing codenames, documents, benchmarks or outputs. They also cannot use the models in real products, because Google can change them or cut access at any time. Once a model launches, testers are encouraged to talk about it freely.

This explains why leaks about test models keep appearing. Partners can't talk, but the models still leave traces. A strong unnamed model on LMArena gets noticed by thousands of voters. Codenames turn up in code and app files. Leakers then join the dots, often weeks before an announcement. A screenshot of the rules page itself is a leak of the same kind.

For regular users, the signal matters more than the codename. Partner testing is usually one of the last steps before a public launch, which is why Lentils80 calls the release very close. If the reports are right, a new Gemini model could arrive soon, with better 3D skills and slower answers.

Treat these details with caution until Google announces Carbon itself.