OpenAI announced a new Ultrafast mode for GPT-5.6 Sol on August 13, 2026, delivering up to 750 output tokens per second, which is approximately 14 times faster than the standard processing tier. The capability is powered by Cerebras Systems, whose wafer-scale chips are currently the fastest commercially available inference hardware.
The announcement is currently in limited preview for select API customers, with no pricing disclosed and no confirmed timeline for general availability.
The Speed Numbers in Context
To understand what 750 tokens per second means in practice, a typical paragraph of text is roughly 150 to 200 tokens. At Ultrafast speeds, GPT-5.6 Sol would produce a 500-word response in under a second, compared to somewhere between 10 and 15 seconds at standard speeds.
For most business tasks, that gap is irrelevant. Nobody is timing how fast a model drafts a contract clause or summarises a report. But for a specific class of applications, particularly those that interact with people in real time or need to process high-frequency events, the speed difference changes what is actually buildable.
OpenAI specifically called out four enterprise use cases for Ultrafast:
- Incident response — security and operations teams need analysis in seconds, not minutes, when systems are failing
- Customer service and support — AI agents handling live chat or voice interactions where lag creates a noticeably poor experience
- Financial market analysis — time-sensitive signal processing where the window closes quickly
- E-commerce — product recommendations, inventory queries, and checkout assistance at scale during peak periods
Intelligence Stays the Same
One critical detail from the announcement: Ultrafast mode runs the same GPT-5.6 Sol model at the same intelligence level as standard mode. This is not a smaller or stripped-down model. It is the same model, faster.
That distinction matters because earlier attempts at “fast AI” often involved running a less capable model. The pitch here is that you get frontier reasoning at speeds previously associated with simpler retrieval or keyword-matching systems.
Cerebras achieves this through its wafer-scale architecture, which solves the memory bandwidth bottleneck that limits inference speed on conventional GPU clusters. Rather than processing tokens across thousands of individual chips coordinated over high-latency interconnects, Cerebras processes them on a single massive chip where memory access is orders of magnitude faster.
What This Means for Business
The enterprises that will care most about Ultrafast in the near term are those with hard real-time requirements, specifically voice AI deployments and live customer interaction systems.
If you are running an AI agent that answers inbound calls, a 10-second response delay is a failed product. A 1-second response is functional. A sub-second response with full GPT-5.6 Sol intelligence changes the calculus for what you can build with voice AI.
For business owners evaluating AI-powered customer service or support infrastructure, the arrival of Ultrafast mode signals that the latency excuse for not deploying frontier AI in real-time contexts is getting harder to sustain. The intelligence is there. The speed is arriving.
A few things worth noting about the current state of this release:
It is still limited preview. Ultrafast is not available to all API customers yet. If your business depends on it, get on the waitlist early and plan your production timeline around access, not the announcement date.
No pricing is public. Speed has historically cost more. Cerebras inference is premium infrastructure, and Ultrafast mode will almost certainly carry a price premium over standard GPT-5.6 Sol. Budget accordingly once pricing is disclosed.
Voice AI is the most immediately obvious application. If you are thinking about deploying voice agents for your business, whether for customer service, internal helpdesks, or automated outbound, this development is directly relevant. The combination of frontier reasoning and real-time response speed removes a major architectural constraint.
The competitive pressure will accelerate. Google, Anthropic, and others will respond. Inference speed is becoming a product dimension the same way model quality has been. Expect similar announcements across the major providers over the next few months.
The broader takeaway is that the hardware and model layers of AI are converging faster than most business planners expected. The ceiling on what AI can do in real-time contexts is rising quickly.
Enterprise DNA builds AI-powered voice employees and agentic systems for businesses ready to move beyond pilots. Book a discovery call to explore what Omni Voice or Omni Ops could do for your operation.