Seven hundred and fifty tokens per second. Let that sink in. OpenAI’s new Ultrafast mode — now in preview on the API — runs GPT-5.6 Sol at speeds that make the already-fast Sol model feel like it’s been given a nitro injection. We’re talking up to 14 times faster than the standard tier, powered by Cerebras and its wafer-scale silicon. This isn’t a minor performance tweak. At this speed, the gap between AI output and human reading comprehension effectively disappears.
Why Speed Matters More Than You Think
There’s a version of this story where 14x faster sounds like a spec sheet flex — impressive on paper, irrelevant in practice. That version is wrong, and here’s why.
Most people interact with AI through chat interfaces where waiting a second or two for a response is perfectly fine. But developers building applications — voice assistants, real-time coding tools, document processing pipelines, agentic workflows that chain dozens of model calls together — live and die by latency. A 7-second wait that’s acceptable in a chat window is catastrophic inside an automated pipeline where that call happens 50 times per task.
GPT-5.6 Sol was already OpenAI’s speed-optimized model, designed for exactly these high-throughput scenarios. The standard API tier was handling it fine. But “fine” doesn’t cut it when competitors are racing to make inference faster and cheaper. Ultrafast is OpenAI’s answer to the question developers have been quietly asking for months: can you push this further?
Apparently yes. Way further.
The timing isn’t accidental either. Enterprises are already building serious production workloads on GPT-5.6 Sol, and as those workloads scale, speed becomes a cost issue as much as a performance issue. Faster inference means fewer compute-hours per task, which means lower bills at scale. OpenAI is solving a real pain point before it becomes a churn risk.
What Ultrafast Actually Is — The Technical Details
The key partnership here is with Cerebras Systems, the chip company that’s been building a very different kind of AI accelerator. Where NVIDIA’s GPUs run inference across thousands of smaller chips connected by high-speed interconnects, Cerebras builds single wafer-scale chips — literally the entire silicon wafer becomes one processor. The result is dramatically lower communication overhead between compute units, which translates directly into faster sequential token generation.
That architectural difference is exactly why Ultrafast can hit 750 output tokens per second. It’s not just better software or smarter batching. It’s fundamentally different hardware optimized for this exact problem.
Here’s what OpenAI has announced for the Ultrafast preview tier:
- Model: GPT-5.6 Sol (the same model, same weights — just faster inference)
- Speed: Up to 750 output tokens per second
- Speed multiplier: Up to 14x faster than the standard API tier
- Infrastructure: Powered by Cerebras wafer-scale processors
- Access: Available now in preview via the OpenAI API
- Tier: Separate service tier — you opt into Ultrafast explicitly
The fact that it’s the same model with the same weights is worth emphasizing. Developers don’t need to re-evaluate output quality, re-tune prompts, or worry about behavioral differences. If your app already runs on GPT-5.6 Sol, switching to Ultrafast is — at least conceptually — a configuration change, not a migration.
Pricing details for the preview period haven’t been fully disclosed, which is typical for OpenAI previews before they move to general availability. Expect a premium over standard Sol pricing, though the compute efficiency gains from Cerebras hardware may keep that premium tighter than you’d guess.
Who Actually Benefits From This?
Real-Time Voice and Conversation Applications
This is the most obvious use case. At 750 tokens per second, a voice AI assistant can generate a complete, nuanced response faster than text-to-speech can consume it. That eliminates one of the main remaining sources of perceptible latency in AI voice products. Companies building voice agents — customer service bots, AI phone systems, accessibility tools — have been working around model speed constraints for years. Ultrafast removes that constraint almost entirely.
Agentic Pipelines and Multi-Step Workflows
This is where I think the real impact lands. Enterprises are already moving from AI chat to AI that actually takes actions — running code, calling APIs, making decisions in sequence. An agentic task that requires 20 model calls at standard speed might take 30-40 seconds end-to-end. At Ultrafast speeds, that same pipeline could complete in 2-3 seconds. That’s not a marginal improvement. That changes what kinds of automated workflows are even worth building.
Document Processing and Batch Analysis
Legal, finance, and compliance teams processing thousands of documents don’t care about individual response latency the way a chatbot user does. They care about throughput — how many documents per hour, per dollar. At 14x the speed, the economics of large-scale document analysis shift meaningfully. Tasks that required overnight batch jobs could run in real time.
Gaming and Interactive Experiences
This one gets overlooked, but at sub-second full-response times, dynamic AI-driven narrative in games becomes genuinely viable. NPC dialogue, procedural storytelling, real-time game master systems — these have all been limited by model speed. Ultrafast opens a door that’s been mostly closed.
The Competitive Picture
OpenAI isn’t the only one chasing inference speed. Groq, which also uses custom silicon (Language Processing Units, or LPUs), has been the go-to option for developers who needed blazing-fast inference on open-weight models like Llama. Groq has been competitive on raw token throughput for smaller models for a while now.
The difference is model quality. Groq’s speed advantage has traditionally come on open-source models that don’t match frontier quality. Ultrafast delivers frontier-level output — GPT-5.6 Sol is a serious model — at speeds that compete with or exceed what Groq offers on lighter models. That’s a meaningful combination that didn’t exist before this announcement.
Anthropic and Google haven’t made comparable speed announcements for their frontier models. Claude’s API is fast but not at this tier. Gemini Ultra has strong throughput through Google’s TPU infrastructure, but Google hasn’t positioned it explicitly as an ultrafast offering with these kinds of benchmark numbers. OpenAI is, at least for now, the only company offering this combination of quality and speed through a standard commercial API tier.
I wouldn’t be surprised if Anthropic announces something similar within the next few months. The pressure to match this is real.
What This Means for You
If you’re a developer already on GPT-5.6 Sol
Sign up for the preview now. Even if you don’t immediately migrate your production workloads, understanding the performance characteristics of Ultrafast while it’s in preview — when you can experiment without full production stakes — is worth the effort. By the time it hits general availability with finalized pricing, you’ll want to already know whether it fits your cost model.
If you’re evaluating AI infrastructure for a new project
Ultrafast changes the calculus for real-time applications. If speed was previously a blocker for a use case you’d considered, it’s worth revisiting that evaluation now. The 750 tokens/sec ceiling is high enough to enable product experiences that simply weren’t practical six months ago.
If you’re building AI-powered products for enterprise customers
Speed and reliability are the two things enterprise buyers ask about most after security. Ultrafast gives you a concrete, impressive answer to the speed question. Pair that with OpenAI’s expanding enterprise security posture and you have a stronger sales story than you did a week ago.
Frequently Asked Questions
What is GPT-5.6 Sol Ultrafast mode?
It’s a new OpenAI API service tier that runs the GPT-5.6 Sol model at up to 750 output tokens per second — up to 14 times faster than the standard API tier. It uses Cerebras wafer-scale hardware to achieve this speed without changing the underlying model.
Is the output quality different from standard GPT-5.6 Sol?
No. Ultrafast uses the same model weights as standard GPT-5.6 Sol — the only difference is inference speed. Developers should expect identical output quality, which means no prompt re-engineering or quality re-evaluation is required when switching tiers.
How does this compare to Groq or other fast inference providers?
Groq has offered high-throughput inference for a while, but primarily on open-weight models like Llama. Ultrafast delivers comparable or higher speeds on a frontier-quality commercial model, which is a different value proposition. For teams that need both speed and top-tier output quality, there’s no direct equivalent right now.
When will Ultrafast be generally available?
It’s currently in preview as of August 13, 2026. OpenAI hasn’t announced a general availability date yet. Preview periods for OpenAI API tiers have historically run anywhere from a few weeks to a few months before broader rollout.
The Cerebras partnership signals something bigger than one fast API tier — it’s OpenAI actively diversifying its inference infrastructure beyond its own data centers and NVIDIA hardware. As frontier models get heavier and demand keeps growing, that infrastructure flexibility is going to matter. Watch this space: the race for inference speed is just getting started, and 750 tokens per second may not be the ceiling for long.
