Google Gemma 4 Under Apache 2.0: 31B Model Matches GPT-4o on Structured Tasks at Zero Per-Token Cost

Google Blog·medium signal

Google's Gemma 4 family (released April 2, Apache 2.0) includes a 31B dense model hitting 85.2% MMLU Pro — competitive with GPT-4o on structured tasks at zero marginal cost once self-hosted. The 26B MoE variant runs on a single GPU with only 4B active parameters per query. Teams locked into proprietary APIs are paying 3-10x more at scale; Midjourney reportedly cut monthly AI spend from $2.1M to under $700K by self-hosting. For SaaS builders, this collapses the economics of adding AI features — a $50/month server can run near-zero-cost inference.