Gemini 4 Carbon is now available
Google is rolling out Gemini 4 Carbon, its most capable model, aimed at hard tasks like multi-step coding, scientific research, and cybersecurity work. It costs more to run than the lighter Gemini 4 Argon, but Google says it's cheaper per task than competing top models on several benchmarks.
- Priced at $2 per million input tokens and $10 per million output tokens, the same introductory rate as Argon, and Google claims more than 2x lower cost per task than the next best external model on CursorBench 4.0, WANDR, and AutomationBench.
- Google previews /swarm in Google Antigravity, which coordinates multiple agents with Carbon handling architecture and auditing while lighter models do parallel work.
- In cybersecurity, Google reports Carbon found 0-day vulnerabilities in two major SaaS providers in under two hours, though these claims come from Google and its partners rather than independent testing.
Google's Gemini 4 Carbon is now generally available through the API in Google AI Studio and Gemini for business, priced the same as its sibling Gemini 4 Argon at an introductory $2 per million input tokens and $10 per million output tokens. Carbon is positioned as the heavier model: Google says Argon remains the right choice for most software engineering and knowledge work, while Carbon is aimed at hard, multi-step problems like complex debugging and open-ended research, using more tokens to explore and verify. Google also previewed /swarm in Google Antigravity, which lets Carbon act as an architect or coordinator for a team of cheaper agents working in parallel, with adversarial verifiers checking each change. Carbon ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, and Google claims more than 2x lower cost per task than the next best external model on those.
The most concrete material is on cybersecurity, which Google first gave to defenders through its Fairwind Program two weeks before this launch. Internal results for Project Fortify, Google's autonomous scanning and triage system, include 100% recall on scanning and 95% precision on triage, and a more than 25% relative recall improvement over Argon on open-source repository testing. Google's Cloud Vulnerability Research team used Carbon to build a proof of concept showing two open-source bugs were exploitable, which led maintainers to reclassify one as critical and ship a patch quickly. Wiz reported that Carbon found 0-day vulnerabilities in two major SaaS providers in under two hours. Most of these are vendor-reported, and the Open Source and Wiz details are either internal or pending disclosure, so they are worth treating as promising rather than independently verified.
On the knowledge work and science side, the headline numbers are a #2 ranking on Artificial Analysis's new Multimodal Knowledge Agent benchmark at more than 6x lower cost per task than the top model, and a top ranking among 23 configurations on Balyasny's financial intelligence benchmark. Axiom Bio says Carbon surfaced toxicity mechanisms it had not found before, and Google researchers used it inside an agentic framework called AlphaProof Code to help formally verify parts of the Signal protocol in Lean, a job they say would normally take expert teams years. Carbon also helped debug lab assay problems and improve chromatography peak deconvolution. The pattern across the post is that Google is leaning hardest on benchmarks it partly controls or co-created and on anecdotes from partners, so the cost-per-task claims in particular deserve a check against your own workloads before you rebuild anything around them.