Bloomberg: Gemini 4 is a very large model, appears affected by benchmaxxing
firstadopter · x · 2026-10-02
Bloomberg reports (cited approvingly by firstadopter) that Gemini 4 is a very large model per a person familiar with its development — big models are expensive to run, potentially pressuring Google's margins. Experts say the model appears affected by "benchmaxxing," the industry tendency to optimize for benchmarks, per two sources. Google thanked its first cohort of cyber defenders and trusted testers, with release starting with paid API customers and Google AI Ultra subscribers.
More from Models
- Compound AI adds Sol 6.1 and argues products should abstract model choice away from users — peterjliu · 2026-10-02
- Benchmarks show Gufo's 70 tok/s claim only holds on repeat-a-word prompts; halogen is faster in practice — brainchillzZ · 2026-10-02
- AI keeps getting cheaper while plans get pricier: the pricing paradox — haider1 · 2026-10-02
- Five AI models try Hemingway's style — GPT, Gemini and Claude end up closer to parody — rvp · 2026-10-02
- Qwen 4 reportedly matches Fable 5 in early samples; 27B due early October — petrusenko_max · 2026-10-02
- ChatGPT Business user finds the seat without custom instructions works far better — BraveBrush8890 · 2026-10-02