Gemini 4 Hydrogen takes #3 on Vending-Bench 2

Gemini 4 Hydrogen placing third on Vending-Bench 2 proves that lightweight, high-throughput models are increasingly viable for complex, multi-turn autonomous agent loops. Long-horizon business simulations demand sustained context discipline rather than raw parameter scale, and Hydrogen's profit-per-dollar ratio fundamentally shifts agent deployment economics.

• –Generating $14,500 in simulated profit places Hydrogen ahead of GPT-6.1 Sol ($13,000) and Claude Sonnet 5.5 ($11,000), trailing only Gemini 4 Carbon and GPT-6 Astra
• –The 365-day simulation stresses multi-step financial management, dynamic pricing, and inventory restocking where earlier models frequently devolved into loops or bankruptcy
• –At $1.50 per million input tokens and 2.5x faster throughput than Flash predecessors, Hydrogen delivers the best cost-to-performance efficiency on the benchmark leaderboard
• –The result demonstrates that configurable thinking levels allow fast models to navigate adversarial supplier delays and customer refund disputes without expensive full-scale frontier inference
• –Strong agentic performance on Vending-Bench indicates that commercial agent builders no longer need to default to top-tier reasoning flagships for autonomous back-office workflows