OpenAI's Ultrafast GPT-5.6 Sol hits 750 tokens/sec, 14x faster
Cerebras-powered and real-time, but gated to a few customers. Nvidia's open Nemotron Lightning is the speed you can actually download today.
Copy markdown
Real-time output: 750 tokens/sec, up to 14x
OpenAI's new Ultrafast mode runs GPT-5.6 Sol at up to 750 output tokens/second — 14x its Standard speed and about 5x Opus 4.8 on Fast. It cleared 2,500 PhD-level questions in 11 hours versus 78+ for rivals, with no quality loss.
The catch: limited preview, no price, no date
Access is gated to a handful of customers behind a sign-up form, and OpenAI says it will 'expand access as capacity grows.' No pricing or GA date yet — so scope your real-time agent ideas now, but don't put production on it this week.
Run it today: Nvidia's open Nemotron 3.5 Lightning
If you want speed you actually control, Nvidia's 30B MoE (just 3B active) is on Hugging Face under the permissive OpenMDW-1.1 license — commercial use OK, fits a single H100. It hits 86% on PinchBench and finishes 10,000 tasks 30% faster than Qwen3.6 35B.
Nvidia also open-sourced a model router: NeMo Switchyard
Switchyard routes each task to the best model in your stack — a self-hostable take on the paid gateways builders lean on. Pair it with Lightning for the cheap-and-fast path and escalate to a frontier model only when a task needs it.