Independent Tests Find Gemini 4 Carbon Near Opus 5.5 in Coding but Slower and Short on Judgment
Google DeepMind's new flagship Gemini 4 Carbon launched just one day after Anthropic's Fable 5.2, and multiple independent reviewers quickly published hands-on tests with strikingly consistent conclusions: coding ability has improved significantly, reaching roughly 90% of Opus 5.5's level, but the model is slower, still lags in judgment and frontend design, and most reviewers ultimately went back to Anthropic's models. Notably, heavy usage burns through quotas extremely fast, raising questions about practicality.
Confirmed
- Indie developer yihuiindie, after two days of heavy use of Carbon (burning 50% of an Ultra 20x account's quota), returned to Claude, saying its coding is 90% of Opus 5.5 but too slow, and criticizing the poor harness experience in Antigravity
- Developer Ben Blaiszik (relayed by iskander) tested Carbon in the Antigravity harness and Argon environment, finding its coding roughly on par with Opus 5.5 but slower, with frontend design still weak; he personally still slightly prefers Anthropic's models
- Every.to got early access from Google (with no editorial interference) and published a Vibe Check hands-on: major coding improvements, but judgment still falls short of Fable 5.2
- Every's user Kieran Klaassen tested Carbon in his self-built Fable ambient music workstation, having it mix a track after rendering 30 seconds of audio, and likewise concluded its judgment doesn't match Fable 5.2
- Every's Mike Taylor compared agents using the same frozen backlog and codebase, letting each model pick tickets it could complete independently and open a PR per ticket: Carbon finished 16 tickets in 4.5 hours, but took far longer than its rivals
Why it matters
- Carbon and Fable 5.2 launched nearly simultaneously, marking a head-on clash between Google and Anthropic in the coding agent race; converging conclusions from multiple independent tests suggest Google is closing in on the leaders in raw coding ability, but still has clear gaps in speed, judgment, and product experience (the Antigravity harness)
2026-11-15 ~ 2026-11-17 · 7 related posts
Primary sources
- [source] Gemini 4 Carbon Finishes ~16 Coding Tickets in 4.5 Hours, but Fable 5.2 Raises the Bar — every · 2026-11-15
- Every's vibe check: Gemini 4 Carbon codes much better but Fable 5.2 raised the bar — every · 2026-11-16
- [source] Every's Vibe Check: Gemini 4 Carbon Codes Better but Judges Worse Than Fable 5.2 — every · 2026-11-16
- [source] Indie dev: Gemini4 Carbon matches 90-95% of Opus 5.5 for coding but is painfully slow — yihui_indie · 2026-11-16
- Carbon Matches Opus 5.5 for Coding but Slower, Says Dev — Though Frontend Design Still Sputters — iskander · 2026-11-16