Google launches Gemini 4 Carbon frontier model, discounted 50% on Vercel AI Gateway

Gemini 4 Carbon arrives

Google has made Gemini 4 Carbon available to developers, enterprises, scientists and consumers, according to the company's announcement. Google describes Carbon as its strongest model to date, built for the kind of work that tests the limits of a model's intelligence: complex multi-step coding and debugging, open-ended research, and security analysis.

The release follows Gemini 4 Argon, and Google is drawing a clear line between the two. Argon remains the better fit for most everyday software engineering and knowledge work, while Carbon spends more tokens exploring a problem, reasoning deeply and running extended verification loops — trading cost for depth on harder tasks. Carbon was previously seeded to cybersecurity defenders through Google's Fairwind Program.

Benchmarks and pricing

Google reports that 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. On the Multimodal Knowledge Agent benchmark, a new evaluation Google built with Artificial Analysis for agents working over dense documents and data, Carbon placed second while averaging more than six times lower cost per task than the leader. Google also says it matched the top-performing model on Hebbia's E2E Financial Agentic evaluation and finished first among 23 frontier model configurations on Balyasny Asset Management's General Financial Intelligence benchmark.

API access via Google AI Studio and Gemini for business is priced at the same introductory rate as Argon: $2 per million input tokens and $10 per million output tokens. The model is also rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search.

Agentic coding and /swarm

On the engineering side, Carbon targets work that spans multiple agents and subsystems, such as multi-package migrations, cross-subsystem root-cause investigations and architectural redesigns. For web development, Google says it can run build-and-verify loops, recover from errors, and check both its code and the rendered interface before delivering a finished application. Those visual and spatial abilities also extend to 3D game development and creative coding.

Alongside the model, Google is previewing /swarm in Antigravity, which expands a single prompt into a coordinated group of agents that can be tracked in real time. Carbon is assigned to the roles requiring the deepest reasoning — architectural design, coordination and auditing — while lighter, faster models work in parallel isolated workspaces and adversarial verifiers test every change.

Security as a headline capability

The most distinctive claims concern cybersecurity. Accessed through the Fairwind Program, Carbon leads the benchmarks Google cites for finding (SEC-bench Pro), exploiting (ExploitGym) and patching (CWE-bench v1) vulnerabilities.

In Project Fortify, Google's autonomous scanning, triage and remediation system, Carbon reportedly hit 100% recall on scanning and 95% precision on triage on an internal benchmark, and improved recall by more than 25% relative to Argon on open-source repository testing. Google's Cloud Vulnerability Research team used the model to demonstrate that two bugs in a popular open-source project were genuinely exploitable, prompting maintainers to reclassify them as critical and ship a patch; details will follow responsible disclosure. Under Wiz's Scan for Good program, the model surfaced 0-day vulnerabilities at two major SaaS providers in under two hours, work Google says would normally take weeks.

Available through Vercel's AI Gateway

Vercel announced that Carbon is also available on its AI Gateway, at 50% off through November 30, 2026. According to Vercel's changelog, the model has a 1M-token context window suited to large codebases and long inputs.

Developers can call it using the model name google/gemini-4-carbon, or wire it into coding agents such as Claude Code, Codex and Cursor by installing the latest Vercel CLI and running its AI Gateway setup command. Vercel says the gateway passes through provider pricing with no markup or platform fee on inference, including bring-your-own-key requests, and adds unified usage and cost tracking, retries, failover, Zero Data Retention support, budgets and routing rules.

Why it matters

Two things stand out. First, Google is deliberately splitting its frontier line — Argon for volume work, Carbon for the hardest reasoning — so model selection becomes a cost-versus-depth decision that teams will have to manage deliberately. Second, pricing is now a competitive weapon: Google pairs its benchmark wins with cost-per-task comparisons, and Vercel's half-price promotion makes experimentation essentially free through the end of November. If the security results hold up outside Google's own evaluations, a model that can find, prove and patch vulnerabilities in hours rather than weeks could materially change how open-source security work gets done.