OpenAI launched Ultrafast on August 13 — a new service tier that runs GPT-5.6 Sol at up to 750 tokens per second, roughly 14x faster than Standard processing. It runs on Cerebras’ wafer-scale hardware, it’s currently in preview, and it’s aimed at workloads where latency is the limiting factor: incident response, real-time customer support, agentic loops, live voice. The intelligence is identical to Standard GPT-5.6 Sol. The speed is not.
Same Model, Different Silicon #
Ultrafast is not a new model. It’s the same GPT-5.6 Sol you’re already using — same weights, same benchmarks, same output quality. What changes is the hardware running it.
Cerebras builds chips differently. Instead of dicing silicon wafers into individual chips and connecting them with bandwidth-constrained interconnects, Cerebras uses an entire wafer as a single processor. Their WSE-3 chip spans 46,000 mm² and packs 44 GB of on-chip SRAM. That’s where the speed comes from.
GPU inference has a fundamental bottleneck: model weights live in HBM (off-chip memory) and must shuttle back and forth to the compute cores on every forward pass. At GPT-5.6 Sol’s scale, that round-trip is expensive. Cerebras eliminates it. Weights stay on-chip. No HBM round-trip. You get 750 tokens per second instead of 53.
OpenAI and Cerebras have a $20B+ multi-year infrastructure deal backing this. This isn’t a proof of concept.
What Changes at 14x Speed #
Speed at this level is a category change, not an upgrade. Here’s what that actually means for builders:
Agentic loops get viable. A five-step agent chain that took 30 seconds at Standard rates finishes in about two seconds at Ultrafast. That’s the difference between a background job and something you can put in front of a user synchronously.
Incident response becomes real-time. OpenAI’s headline use case is analyzing logs, recent code diffs, and engineer reports while the outage is still live. At 53 tokens/sec, the analysis is done after the post-mortem. At 750 tokens/sec, it’s useful during triage.
Voice AI becomes frontier-quality. Conversational AI requires response latency under 300ms to feel natural. Standard GPT-5.6 Sol doesn’t get there on complex prompts. Ultrafast does. This is the first time you can put a frontier reasoning model in a live voice loop without workarounds.
Live support stops faking it. Real-time customer support has always involved the model racing the customer’s patience. Multi-step reasoning that requires tool calls and context lookup now completes before the user notices they’re waiting.
OpenAI also calls out financial analysis and e-commerce — any domain where the data has a short freshness window and the answer needs to arrive while conditions still hold.
When to Use Ultrafast vs. Standard #
Not everything needs Ultrafast. The useful question is: will the user or system notice a two-second delay? If the answer is no — batch jobs, overnight pipelines, document processing queues — Standard is cheaper and plenty fast enough.
Ultrafast makes sense when you’re building one of the following:
- Live user-facing interfaces where latency is UX
- Real-time voice or streaming conversation
- Agentic chains where intermediate step latency accumulates
- Operational systems that need LLM reasoning at machine speed
- Security or incident response tooling where delay is cost
If you’re routing the same workload through both tiers, design your pipeline to tag latency-sensitive steps explicitly. Don’t pay Ultrafast rates for steps that don’t need it.
How to Get Access #
Ultrafast is invite-only for now. OpenAI is running a waitlist at openai.com/form/ultrafast. They ask for your workload type, latency requirements, and expected usage volume. The more specific you are about a real production use case, the better your position in queue. “We’re curious” will not be prioritized.
Pricing hasn’t been announced. Expect a premium over Standard rates at GA — the infrastructure cost justifies it. Budget accordingly before locking your architecture to it.
The Bigger Picture #
The OpenAI-Cerebras deal isn’t just about making ChatGPT faster. It signals that OpenAI is diversifying its compute stack beyond a single hardware provider. Training stays on Nvidia. Inference — especially latency-sensitive inference — moves to specialized silicon where the architecture fits the workload.
Expect Anthropic and Google to respond. Fast inference is no longer a niche conversation. The compute tier you choose for production inference is becoming a first-class API decision, not an afterthought. Read OpenAI’s full announcement and start thinking about it before your architecture is locked.