While everyone's geeking out over Grok4's insane physics sims and Kimi K2's 1T OS bombshell (crushing coding benchmarks for pennies), the real AI drama is in the pricing shadows. After my LLM Selector post blew up here, I kept getting DMs asking "but which provider should I actually use?" So I dove deep into 439 models across 63 providers.
What I found? some interesting insights:
1. huge markup on identical models
Take DeepSeek R1 0528 (quality 68 from Artificial analysis bench, beats many flagships):
Completely free on Google Vertex and CentML (decent speeds too, 121 tok/s and 87 tok/s).
But jumps to $0.91 on Deepinfra, $4.25 on Fireworks Fast, and a whopping $5.50 on SambaNova, for the exact same model (ofc with speed differences).
Arbitrage alert: Why pay infinite markup when free tiers deliver the goods for experimentation or bulk runs?
2. Latency goldmines hiding in plain sight
Sub millisecond responses aren't just for premium setups:
Nebius Base crushes it with DeepSeek R1 at 0.61ms latency for $1.00/1M (103 tok/s) and Qwen3 235B at 0.56ms for $0.30/1M (50 tok/s).
Groq takes it further with models like Qwen3 32B at 0.14ms for $0.36/1M (627 tok/s).
Arbitrage alert: These blow away slower "enterprise" options costing 10x more, ideal for real-time apps