Image: Google
Google opened Gemini 4 Carbon to everyone this week, two weeks after handing it to a small group of security defenders through its Fairwind program. The company calls it its strongest model. The detail that matters most for anyone planning to build on it is the invoice. Carbon sits on introductory pricing until November 30. After that it goes from $2 per million input tokens and $10 per million output to $4 and $20.
That footnote is easy to miss, and it changes the maths on anything you test this month. Google charges the same introductory rate it gave Gemini 4 Argon, so the model costs the same as Argon today and twice as much in three weeks.
Gemini 4 Carbon is now available. It's our strongest model, built to dig deeper, handle long-horizon tasks, and multi-step reasoning. If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems.
— Google (@Google) November 10, 2026
The second thing worth knowing is that Carbon does not replace Argon. Google's own guidance says Argon stays the right pick for most software engineering and knowledge work, and Carbon is for jobs that need deeper reasoning and more tokens spent on exploration and verification loops. This is a second, more expensive tier, not an upgrade path. When Google announced Carbon, it kept it away from the public for exactly that reason. Now the gate is money rather than access.
On benchmarks, Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than twice lower cost per task than the next best external model. It ranks second on a new Multimodal Knowledge Agent test that Google built with Artificial Analysis, at more than six times lower average cost per task than the top model. Those are benchmarks Google chose and helped design, and the cost figures compare spend per task rather than token prices, which flatters a model that gets the job done in fewer attempts.
The feature Google is pushing hardest is /swarm in Antigravity, now in preview. One prompt becomes a team of agents working in parallel, with Carbon given the senior jobs like architectural design, coordination and auditing, while lighter models build. The demos read like a flex. A Three.js reconstruction of the Eiffel Tower between 1887 and 1889 from historical photos and engineering plates. A 906,000-triangle model of Google's Platform 37 office built from planning PDFs over ten overnight loops. An open-world game where you explore the Mountain View campus as a pelican riding a GBike.
The cyber results are where Google can point hardest. Carbon leads the benchmarks for finding, exploiting and patching vulnerabilities when accessed through Fairwind. Inside Google, Project Fortify hit 100% recall on scanning and 95% precision on triage on an internal benchmark, and showed more than a 25% relative improvement over Argon's recall on open-source repositories. Wiz used Carbon to find zero-day bugs in two major SaaS providers in under two hours, a job that would normally take weeks.
The safety numbers moved too. Google says Carbon blocked 99% of attack trajectories from its own red-teaming agent and cut the per-attempt breach rate to 0.8% from 3.1% on Argon. External testers at Trajectory Labs found no new universal jailbreaks across 132 hours and more than 13,000 requests. On a 375-task agentic honesty benchmark it passed 92%, against 70% for Argon, declining to fill in missing experimental data 98% of the time and flagging flawed methods 68% of the time.
Google also repeats that Carbon hit a 0% prompt injection success rate on the Gray Swan benchmark, which it calls a first. It then adds, more honestly, that the model still needs work in that area and that sophisticated attacks can still get through in some distributions. That second sentence matters more than the first number.
The research section is where the model looks least like a product demo. Google researchers used Carbon inside AlphaProof Code to generate millions of lines of Lean, the proof language, for a formal argument about the security properties of Signal's cryptographic protocol. They say the work took weeks rather than the years it normally takes expert teams. Another group used it to find remote protein homologs for seven low-homology families from the CASP17 blind test set, raising the number retrieved by 40-fold on average, which feeds structure prediction models without retraining them.
If you are still on Gemini 2.5, you have a separate deadline. 2.5 Pro and 2.5 Flash retire on December 1, the same day Carbon's price doubles. Two dates, one migration plan, and three weeks to run it.
Comments