Start building with Gemini 4 Carbon: our strongest model for complex, long-horizon tasks. It’s designed to explore problems thoroughly and run extended verification loops across software engineering, cybersecurity, knowledge work, and science.

From synthesizing dense multimodal files to building interactive 3D web experiences, or even uncovering software vulnerabilities, Carbon helps solve problems that require sustained reasoning. Now rolling out in: 🔘 The @GeminiApp and AI Mode in Search for Google AI Ultra subscribers. 🔘 Gemini 4 Carbon is available via the API in @GoogleAIStudio and Gemini for business for all paying developers and enterprises 🔘 Gemini for business, CodeMender, and Google Antigravity for enterprise use

With @ArtificialAnlys, we built MKA (Multimodal Knowledge Agent): a benchmark for how well agents handle visually complex tasks. Everyday professional work often uses more than just text – from cross-referencing blueprints to reading scientific charts. This is not covered by existing evaluations. Gemini 4 Carbon ranks #2, at 6x lower average cost per task than the top model. →

When pro users 🙄

Una IA que verifica su propio trabajo ya supera a medio LinkedIn

Added it to our LLM to 3D benchmark, it ranks #1. You can find some matchups below.

how long can it run before it stops to ask "should i continue"? fable went 11 hours straight on this for me

Extended verification loops sound great, until the loop is still going and dinner's ready lol

Extended verification loops is the part that matters most to me. A model that checks its own work before answering saves hours of fixing. How long can it run on one task before it needs a human? 👀

how long can one Carbon run go before it stops ?

I wonder what the RL tasks will look like for the next iteration.

gonna throw my messiest side project at it tonight

Im still waiting on the bank to approve my loan, so i can afford it.

The verification loop is what I would actually use it for. Long tasks go wrong in the middle, and a model that keeps checking its own work beats one that just runs longer.

66 tool calls across cad, plates, photos, and datum shows real orchestration

Extended verification loops drive up token usage and latency fast. The key tradeoff will be whether the pass rate gains justify the extra compute cost on everyday repo runs.

verification loops break when the same run writes the summary and grades it. the check has to run somewhere that never saw the plan.

I'd hesitate on the verification loop. I keep changing the spec mid-run, so what it's checking usually isn't what I want anymore.

Hmm, the extended verification loops part caught my eye. I still run my own checks by hand after the model finishes, and I keep catching the small misses it walks past.

Longer runs are fine. Just let me see what it's checking.

Google says Carbon helped AlphaProof Code prove key security properties of code behind Signal's protocol. More interesting to me than another leaderboard: what was proved, under which assumptions, and what remains outside the guarantee? A proof is only as broad as its scope.