Google Just Announced Gemini 4 Carbon (It Cooks)

Summary

Google announced Gemini 4 Carbon, a new frontier model built on Gemini 4 Argon, which was released two weeks earlier. Presenter Paul J. Lipsky describes Carbon as a substantial upgrade, while noting that the performance figures are Google’s own benchmark claims.

Capabilities and performance

• Long-horizon tasks: Google says Carbon can work on complex engineering tasks for days and persist through dead ends. Lipsky highlights this as a potential improvement for AI agents, which can stall or give up on difficult tasks.
• Cybersecurity: Carbon is designed to find software vulnerabilities, verify them, and help patch them. Google says it led in all three of its security tests. On Grey Swan’s prompt-injection benchmark, Google reported a 0% attack success rate.
• Benchmarks: Carbon reportedly beat competing models on 16 of the 18 benchmarks Google shared and surpassed Argon on all of them. It reportedly fell behind on one coding benchmark and one computer-use test. On WebDev Arena, it ranked fourth.
• Reliability: Google says Carbon makes about half as many unsupported claims as Argon. It also scored 0% on a coding benchmark measuring whether a model cheats to appear successful.

Availability and pricing

Carbon is not generally available yet. Google plans to provide access first to a small group of cyber defenders, then to trusted partners, before opening it more broadly. Lipsky speculates that public access could follow a timeline similar to Argon’s, but Google has not announced a date.

Developer pricing is expected to start at Argon’s introductory rate of $2 per million input tokens and $10 per million output tokens, then double to $4 and $20, respectively.

Open questions and coverage

Google has not explained how Argon and Carbon fit into a broader model-tier structure. Lipsky plans to publish a hands-on comparison once Carbon becomes available.

This is an announcement analysis, not a hands-on review or tutorial. Benchmark and capability claims are attributed to Google; the prompt-injection benchmark is identified as Grey Swan’s.