Gemini 4 Still Isn't What I Need (Yet)

Gemini 4 Carbon cuts the false-claim rate of its predecessor roughly in half, but it is too slow, too costly, and too error-prone to replace Opus for everyday coding. Its clearest strength is security work, which rests on one developer's own tests rather than independent benchmarks. [04:38]

Where Carbon fits in practice

1. Keep Opus as the default model for everyday coding, since Opus finishes tasks in minutes where Carbon takes twenty or more. [26:00]
2. Turn to Carbon for security review and incident response, especially when another model refuses or waters down the work. [24:14]
3. Update T3 Code to the release that includes the security fixes for the two reported vulnerabilities. [19:12]
4. Avoid Carbon for code review, codebase analysis, and 3D work, where it invents details or produces broken output. [15:20] [28:34]
5. Verify token use and cache behavior before picking Carbon for cost-sensitive work, since the introductory price rises sharply. [09:04]

Topics behind the verdict

Hallucination rate — Across the creator's evaluations, false claims dropped from 36% with Argan to 18% with Carbon. In a Rust architecture review, Carbon still produced 23 false claims out of 79, while most other models produced zero to one. [04:38]

Response speed — Average task time on low reasoning was 22 minutes and on high was 27, against roughly 3 minutes for Opus at low effort. Reasoning settings barely changed the delay, so the slowness is structural. [05:30]

Pricing and caching — The introductory price of $2 per million input tokens and $10 per million output tokens rises to $4 and $20 on December 1st. Gemini caches inconsistently, with a 20% miss chance within five seconds, which can add as much as 50% to requests. [06:40]

Cyber security capability — Google reports Carbon leads on several security benchmarks and found bugs similar in severity to PHP-to-shell flaws in open source projects. The creator confirmed its value by getting a usable fix for T3 Code's vulnerability that other models would not provide. [14:08]