Gemini 4 still fabricates findings in code reviews, side-by-side test finds
qu1etus · reddit · 2026-10-14
A developer who published a paper on multi-LLM code reviews (and abandoned Gemini over hallucinations) tested Gemini 4 (high, via antigravity) against gpt-6.1-sol (Codex CLI) and Opus 5.5 (Claude Code CLI) on the same review.
Result: Gemini 4's three real findings merely duplicated the other models', and its census numbers plus one classifier claim were fabricated — the same pattern that pushed him away from Gemini 3.1 earlier this year. He's disappointed the issue apparently wasn't fixed in the new model.
More from coding & agent
- Ethan Mollick's new book charts 30 months of AI progress, from cyborg to conjurer work — rseroter · 2026-10-23
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Jensen Huang details Vera Rubin agentic stack: NVL72, MGX, CMX and RTX — rohanpaul_ai · 2026-10-23
- Building an autonomous SEO agent in Hermes with three skills: programmatic pages, Keywords Everywhere, DataForSEO — EXM7777 · 2026-10-23
- Building a sports betting AI agent with onchain execution and human approval by default: architecture and lessons — Dry-Comfortable-2678 · 2026-10-23
- Capy 0.5.0 lets Claude Pro/Max subscribers reuse their plan quota, adds OIDC cloud access — garrytan · 2026-10-23