Bloomberg: Google employees worry Gemini 4 coding underperforms; insiders call report detached from reality
xennygrimmato_ · x · 2026-10-02
Per Bloomberg, some Google employees with early access say Gemini 4's strong benchmarks aren't translating to real-world performance, especially on coding and front-end design, and the model may be expensive to run; Google disputes this and says consensus is the model is frontier-level, having also scrapped the planned Gemini 3.5 Pro relaunch. Internal voices like SicongJiang25 push back hard, calling the report detached from reality and citing overwhelmingly positive internal reception.
More from Models
- Leaked-looking model list teases Opus 5.5, Fable 5.1, Sol 6.1 and more — sloppenheimer · 2026-10-02
- JevBench v1.5.4 ditches cost-weighted scoring; Original Jev returns to top of leaderboard — airesearch12 · 2026-10-02
- Dan Shipper says models fail at impersonating him because he agrees only 34% of the time — danshipper · 2026-10-02
- Blogger teases Gemini 4 outputs, says "I'm impressed" — iruletheworldmo · 2026-10-02
- Opus 5.5 high vs GPT-6.1 Sol max: AA data shows Sol's value edge is mostly latency, not intelligence — Wsz2020 · 2026-10-02
- Claude Sonnet 3.7 "returns to its room" and finds notes left in May — repligate · 2026-10-02