OpenAI supercharges GPT-5.6 Sol with 750 tokens-per-second Ultrafast tier

For years, anyone building on top of large language models had to make an annoying compromise: pick a flagship model for real reasoning and wait for it to trickle out words, or drop down to a mini model if you needed real-time responses.

OpenAI set out to kill that tradeoff. The company rolled out Ultrafast, a new API service tier designed to run its top-tier GPT-5.6 Sol at up to 750 output tokens per second. That is roughly 14 times faster than standard processing, turning what used to be a coffee break into near-instant output.

What 750 tokens per second actually looks like

To put that speed into perspective, standard flagship runs of Grok, Claude, and earlier GPT setups typically hover around 60-something tokens per second. GLM 5.2 sits around 111, and even speed demons like Gemini 3.7 Flash clock in around 340 tokens per second. In other words, Ultrafast is more than double the speed of the fastest lightweight frontier competition.

While rivals like Anthropic offer faster tiers like Claude fast mode, as TechCrunch noted, none are operating anywhere close to this tier. In side-by-side demos, standard GPT-5.6 Sol is still generating its first few lines while Ultrafast has already dumped a full page of structured technical analysis and wrapped up the job.

Behind this leap is OpenAI's deepening infrastructure partnership with Cerebras. By leveraging wafer-scale chip architectures designed specifically to eliminate memory bandwidth choke points during inference, they managed to push frontier intelligence through the wire without downscaling the weights.

Where every second is money

Cutting latency by 14x isn't just about finishing ChatGPT prompts faster; it shifts entirely offline batch jobs into live interactive loops.

OpenAI is testing the tier with a select group of pilot customers, including quant powerhouse Jane Street, along with Podium, Basis, and Rogo. The key use cases are places where downtime or delays destroy business value. In critical incident response, an ongoing outage gives engineers mere minutes to comb through logs, traces, and fresh commits; Ultrafast lets Sol digest that firehose in real time while human operators call the shots.

Other targets include real-time financial risk detection, live voice assistance that doesn't stumble over multi-step tool calls, e-commerce checkout recovery before a user abandons a cart, and code research loops. Inside OpenAI, researchers report replacing overnight batch experiments with rapid-fire iterations within a single working afternoon.

Right now, access remains locked behind a limited preview, primarily because capacity on specialized wafer hardware is still tight. OpenAI plans to expand availability as infrastructure scales up.

Why it matters

We see Ultrafast as an important shift in AI competition: the frontier is no longer just about raising benchmark IQ, but maximizing useful cognitive work per second.

For enterprise developers, this solves a brutal deployment roadblock. Building complex agents or voice pipelines previously forced teams into lightweight, error-prone models just to keep latency acceptable. If frontier reasoning can stream at 750 tokens per second, high-stakes tasks like automated debugging and real-time financial arbitration suddenly become viable production workflows instead of fragile prototypes.

For the broader industry, it is also a quiet shot across the bow of traditional GPU inference. While Nvidia-based clusters excel at massive batch processing, Cerebras's specialized wafer-scale approach is proving that extreme low-latency single-stream inference requires fundamentally different hardware. By proving this out at production scale, OpenAI is actively diversifying away from conventional silicon monopolies.

The catch, of course, is economics and scale. Wafer-scale computing is notoriously expensive and difficult to manufacture in volume. If Ultrafast commands a prohibitive premium or remains capacity-constrained, it will stay an exclusive perk for hedge funds and enterprise giants. But as a technical proof of concept, it proves that the speed-versus-smarts compromise is finally coming to an end.