Cerebras Just Made GPT-5.6 Sol Answer at 750 Tokens a Second. Here's What 14x Speed Is Actually Worth to a Solo App
Cerebras Just Made GPT-5.6 Sol Answer at 750 Tokens a Second. Here's What 14x Speed Is Actually Worth to a Solo App
On August 13, OpenAI and Cerebras announced Ultrafast, a new service tier that runs GPT-5.6 Sol at up to 750 output tokens per second, up to 14 times faster than standard processing on the same model. No smaller model, no quality tradeoff, same GPT-5.6 Sol intelligence, just Cerebras' Wafer-Scale Engine hardware doing the generation instead of standard GPU inference. It's in limited preview to a select group of OpenAI customers, no public pricing, no self-serve signup beyond a waitlist form. If you build anything where a user sits and watches the response stream in, this is worth twenty minutes of your attention. If you don't, it probably isn't, and figuring out which camp you're in is the actual point of this post.
What Ultrafast mode actually is
Standard GPT-5.6 Sol runs at roughly 53 output tokens per second based on the math in OpenAI's own announcement (750 divided by the stated 14x multiplier). Ultrafast mode, powered by Cerebras, pushes that to up to 750 tokens per second. Cerebras' side of the announcement explains the mechanism: fast inference is a data-movement problem, and on GPUs, generating each token means shuttling model weights between on-chip memory and off-chip storage over and over. Cerebras' wafer-scale chip packs 44 GB of SRAM directly onto the chip, so weights stay put and tokens flow through the pipelined layers without that back-and-forth. It's the same architecture Cerebras has sold as a speed play for open models like Llama and GPT-OSS for a couple of years now. What's new is OpenAI putting its actual flagship model, GPT-5.6 Sol, behind it.
Cerebras ran its own benchmark to make the point concrete: on Humanity's Last Exam, a 2,500-question benchmark needing PhD-level reasoning across chemistry, economics, and literature, GPT-5.6 Sol on Ultrafast finished all 2,500 questions in 11 hours and 11 minutes. Cerebras says Claude Fable 5 took 78 hours and 27 minutes to do the same run. Take that with the grain of salt it deserves, it's Cerebras' own benchmark on its own hardware, and "comparable accuracy" is doing some work in that sentence. But the direction is real: a task that used to take three days of continuous compute now finishes in one working day.
Availability is the part that matters more than the benchmark, honestly. This is not a toggle in your OpenAI dashboard. It's a limited preview running with an initial group of companies (OpenAI names Jane Street, Podium, Basis, and Rogo as early testers), and OpenAI says access will expand "as capacity grows," with no date attached. If you're not already an OpenAI enterprise customer with a relationship to lean on, you're on the same waitlist as everyone else right now.
Where 14x speed actually changes a product
The use case that keeps showing up in both companies' announcements is voice. Courtland Lykins, who leads voice AI product at Podium, put it plainly in OpenAI's writeup: "For us the Ultrafast has been invaluable in our voice stack. The speed completely changes the call experience for the more complex work." That tracks. Voice has a hard physical constraint that text chat doesn't: a person on the other end of a phone call notices a 2-second pause in a way they'd tolerate in a chat window without thinking about it. If your product is a voice agent handling anything more complex than "what are your hours," standard GPT-5.6 Sol's roughly 53 tokens a second is audibly, physically slow. 750 tokens a second is the difference between a call that sounds like a real conversation and one that sounds like a hold-music interstitial with intelligence bolted on.
The other place this genuinely matters is live coding assistants and anything where a human is watching the cursor move, waiting to intervene. Mitch Troyanovsky at Basis said something in the OpenAI post worth sitting with: "the barrier to truly fast products is not just tokens per second, but also model intelligence, and Ultrafast combines both." That's the actual insight. Groq and Cerebras have both been fast for years running smaller open models. What's new is getting that speed without dropping down to a weaker model to get it. If your bottleneck has genuinely been "the model is smart enough but too slow to keep the user engaged," this closes that gap for the first time on a frontier model.
Where 14x speed is money spent on nothing: batch jobs, anything already async, anything running overnight while nobody's watching. OpenAI's own internal example makes the point well. Its research team describes using Ultrafast to compress a workflow that used to be "launch a batch of experiments overnight, review results in the morning" into something that supports multiple iterations during the workday. That's a real win, but notice what it's not: it's not making the overnight batch job faster in a way anyone needed. It's turning an async workflow synchronous because the model is now fast enough for a human to sit and watch it work. If nobody's watching your job run right now, Ultrafast just buys you a bigger bill for tokens nobody's staring at.
The pricing question nobody's answered
Here's the part every solo builder needs to sit with before getting excited: OpenAI has not published Ultrafast pricing anywhere in the announcement. Not a per-token rate, not a multiplier over standard, nothing. That's unusual for OpenAI, which typically ships pricing right alongside a model or tier launch, and it tells me the economics aren't settled yet, probably because Cerebras' wafer-scale capacity is a real constraint on how much of this OpenAI can sell at any price.
History is not encouraging here. Groq, the other major low-latency inference provider, doesn't run a separate "fast mode" surcharge, but its speed is baked into infrastructure costs that show up as flatly higher per-token rates on some models compared to commodity GPU providers. Cerebras' own existing dedicated-capacity contracts for open models reportedly run in the $0.25 to $0.60 per million token range depending on volume, a real premium over its cheapest self-serve tier. Cerebras also sells flat-rate coding plans today, $50 a month for 24 million tokens a day, $200 a month for 120 million tokens a day, on top of open models, not GPT-5.6 Sol. None of those numbers translate directly to what Ultrafast will cost on a frontier OpenAI model, but the pattern across the entire fast-inference market is the same: speed has never been free, and providers that sell it as a distinct product tier price it as one.
If Ultrafast lands anywhere close to a meaningful markup over standard GPT-5.6 Sol pricing, which is already well above OpenAI's smaller models, that's a real cost decision for anyone not running on enterprise-scale volume discounts. A solo operator burning through a few million tokens a month on a synchronous product might find the latency win worth a meaningfully bigger bill. A solo operator running mostly batch or async workloads would be paying a premium for a capability they can't use.
What I'd actually do
If you're building or running a voice product, a live pair-programming assistant, or anything where a human is watching a response stream character by character and the wait itself is costing you engagement or conversions, put your name on the waitlist now. Being in the first wave when access expands, before your competitors are, is worth the five minutes it takes to sign up, even with zero pricing information in hand.
For everyone else, and that's most solo AI products, I'd wait. Wait for public pricing. Wait for the preview to open past the current hand-picked customer list. And before you do either, actually check whether token generation speed is your real bottleneck. For a lot of products I've looked at, the perceived slowness isn't the model generating tokens too slowly, it's a chain of API calls, a vector search, a tool call round-trip, or a UI that doesn't stream partial output at all. Speeding up the one link in that chain that was never the slow one doesn't fix the spinner problem. It just makes you feel like you did something.
My honest counter-take: I could be underselling this. If Ultrafast pricing comes in cheaper than expected, maybe because Cerebras is subsidizing adoption early the way inference providers often do, or if it turns out solo operators get reasonable access faster than "as capacity grows" implies, the calculus flips fast. A synchronous product that's been artificially constrained to feel less capable than it actually is, because nobody could ship a frontier model at conversational speed, gets meaningfully better overnight, and the Podium and Basis quotes read like genuine product changes, not marketing copy. I'd rather be the person who signed up for the waitlist and didn't need it than the one still filing the paperwork when access opens.
Author
Lukas
@lukcombinator