OpenAI just flipped the speed switch on GPT-5.6 Sol, and the numbers are frankly absurd. A new “Ultrafast” mode, previewing now through the API, pushes output to 750 tokens per second. That’s roughly 14x faster than the model’s standard throughput, and it’s not some lightweight distillation trick — this is the full flagship reasoning model running on Cerebras hardware, the same company that inked a ten-billion-dollar partnership with OpenAI earlier this year.
Don’t expect to toggle this on tomorrow. Access is limited to select customers for now, with a signup form for everyone else and a promise of gradual expansion as capacity comes online. The pitch from OpenAI is that Ultrafast finally closes the gap between small-model speed and large-model reasoning, delivering what the company calls “more useful work per second.” Internally, OpenAI says it’s already using this for incident response — engineers get logs, code changes, and reports analyzed while an outage is still unfolding, not after the postmortem meeting.
The use cases OpenAI is floating span finance, support, e-commerce, and research. Flag suspicious transactions while market conditions are still shifting. Resolve multi-step customer inquiries before the caller loses patience. Answer product questions, check inventory, and personalize recommendations before a hesitant buyer abandons their cart. The research angle is maybe the most interesting: overnight batch jobs become interactive sessions where you test an idea, see results, tweak, and rerun without losing your train of thought.
What’s really happening here is pricing architecture. OpenAI already sells a “Fast Mode” at roughly double the price for 2.5x speed. Ultrafast adds a third tier, faster and almost certainly pricier. It’s the AWS playbook applied to inference — charge a premium for the same model at higher throughput, and capture a slice of whatever revenue gains that speed unlocks for customers. Whether 14x speed justifies the coming markup depends entirely on what you’re doing with it. For real-time incident response, probably yes. For batch processing where latency doesn’t matter, you’re better off with the slow lane.
