Gemini 4 Carbon Is Now on Multos — Google's Best Model, and It Costs Less Than Opus
59 on the Artificial Analysis Intelligence Index. First place, ahead of Claude Opus 5.5 and GPT-6 Astra. It also finishes benchmark tasks in about 123,000 tokens — roughly half what Sonnet 5.5 needs. Carbon is live on Multos Elite today.
Gemini 4 Carbon is Google's strongest model and it is now on Multos Elite. It scores 59 on Artificial Analysis Intelligence Index v5, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. But the number that matters more for people building on Multos is this: it finishes benchmark tasks in about 123,000 tokens, roughly half what Claude Sonnet 5.5 needs. Fewer tokens per task means fewer turns before you hit a wall, and more of your monthly budget left for the next thing you build.
What People Actually Build With It
Carbon earns its place on builds where the hard part is reasoning, not typing. A security review app that scans a repo for vulnerabilities before deploy. An architecture planner that reads an existing codebase and proposes a migration path. A debugging assistant that holds a hypothesis across twenty files and tests it. A research tool that pulls apart a dense document set and answers questions with citations. A compliance checker that reads policy against implementation. What these share is that a wrong answer is expensive and a shallow answer is useless. Carbon is overkill for a landing page, a CRUD form, or a dashboard wired to an API. Those are real builds, and they do not need this model.
The Cybersecurity Numbers Are the Real Story
Google built Carbon with cyber defense in mind and the benchmarks show it. On ExploitGym, which tests finding and exploiting vulnerabilities, Carbon scores 54 against 42 for GPT-6 Astra and 34 for Opus 5.5. It leads SecBench Pro on vulnerability discovery and scores 71 on CWEBench for patching known weaknesses. On SWE-bench it posts 9.6 against Argon's 4.3, and on Terminal-Bench 4.0 it hits 65 versus 59 for both Argon and Astra. If you are building anything that touches security tooling, this is the model with the widest margin over the field. That margin is also why Google staged the rollout through cyber defense partners rather than dropping it into general availability.
Where It Loses
Carbon is not first at everything, and the honest version of this article says so. On OSWorld, which measures controlling a computer across desktop apps, Opus 5.5 leads and Carbon sits second at 81.5. On cost per completed task, OpenAI's GPT-6.1 Sol family is still more efficient, largely because Sol Max uses around 68,000 tokens per task against Carbon's 123,000. Carbon is the highest-scoring model, not the cheapest way to finish a job. If your build is mostly mechanical steps with occasional hard reasoning, a cheaper model plus Auto mode will cost less and finish about as well.
The Alignment Work Nobody Talks About
Google reported specific fixes alongside the launch that matter if you are shipping to real users. Successful cheating on Terminal-Bench 4.0 dropped to 0%, with cheating attempts down to 4%. Unsupported or fabricated claims were roughly halved compared to Argon. For anyone building an app where the model's output goes in front of a customer, that second number is the one to care about. A model that invents a plausible-sounding API or a citation that does not exist is a support burden. Carbon is meaningfully less likely to do that than its predecessor.
What It Costs on Multos
Carbon is an Elite model. Google prices it at $2 per million input and $10 per million output through November 30, 2026, moving to $4 and $20 after that, with cached input at a 95% discount. On Multos that works out to $0.885 per million blended at our standard cache assumption, which is a 1.77x token multiplier on Elite. The introductory rate is the reason that number looks low. When Google raises it in December, the multiplier goes up with it, and we will update it then rather than quietly absorbing the difference.
How to Decide
Reach for Carbon when the build's value depends on getting a hard answer right: security analysis, architecture, long debugging, legal or research work where it ranks first of 87 on Legal Research Bench at 55.77%. Skip it for routine edits, file reads, and the small structured steps that make up most of an agent's token spend. The practical setup for most people is Auto mode, which routes cheap steps to cheap models and reserves Carbon for the turns that need it, with manual selection in the dropdown when you know a task is hard.
Ready to Get Started?
Gemini 4 Carbon is live on Multos Elite. 59 on the Intelligence Index, first place ahead of Opus 5.5 and GPT-6 Astra, at $0.885 per million blended. Open the model dropdown, find it under Google, and point it at the build where a wrong answer actually costs you something.
Start Building FreeRelated Articles
One-Shot Website Generation: Why Most AI Builders Can't Do It (and How Multos Does)
Most AI builders need 10-30 prompts to reach a finished website — and every prompt costs credits. Multos generates a complete, designed, deployed site in one shot. Here is the engineering reason one-shot works, and why competitors structurally can't match it.
Read more →GPT-5.6 Luna Price Cut 80% — Now Available on Free & Lite
OpenAI slashed GPT-5.6 Luna by 80% overnight — from $1/$6 to $0.20/$1.20 per million tokens. We're passing it straight through: Luna is now available on every plan including Free and Lite.
Read more →Claude Opus 5 & Sonnet 5 Now Available: Near-Frontier Intelligence at Every Tier
Anthropic's Claude Opus 5 and Claude Sonnet 5 are now live on Multos. Opus 5 delivers near-Fable-5 intelligence at half the price with 79.2% on SWE-bench Pro, while Sonnet 5 brings Opus-class agentic capabilities to Starter-tier users at $3/$15 per million tokens.
Read more →