Gemini 4 Carbon Finishes ~16 Coding Tickets in 4.5 Hours, but Fable 5.2 Raises the Bar
every · x · 2026-11-15
Every's vibe check of Gemini 4 Carbon (early access from Google, no editorial input) used Mike Taylor's frozen-backlog test: each agent picks tickets it can complete unaided and opens a separate PR per ticket.
- Other agents finished 8–9 tickets in about an hour; Carbon completed 15–16 before being stopped at 4.5 hours — more output, far more time.
- Carbon is a much stronger coder than Gemini 4 Argon, makes better 3D objects, and can run ambitious builds for hours with little supervision.
- The catch is judgment: Anthropic's Fable 5.2 (released a day earlier) showed an agent can work long hours while choosing the right work, respecting scope, and knowing when it's done. Carbon narrows the gap, but Fable has reset expectations for what a flagship should deliver.
- Testers disagree on how much supervision Carbon still needs.
More from coding & agent
- Indie dev shows off T3 Code dashboard with 452 settled agent threads — yihui_indie · 2026-11-15
- Amazon planning World Product Graph to index the web's products for AI shopping agents — jomic01 · 2026-11-15
- Google Labs open-sources 15 AI Design Skills that turn prompts into React UI, 9k GitHub stars — tokumin · 2026-11-15
- 20 GitHub repos turn Fable 5.2 into a full motion studio — 3D, FX, sound, export — PrajwalTomar_ · 2026-11-15
- 79k-Star AI Engineering Course Ships Its Own Agent Tutor for Claude Code and Cursor — ghumare64 · 2026-11-15
- Flappy Bird built with Hala Play + Claude: all art and audio AI-generated — aziz4ai · 2026-11-15