Published: 2026-08-15 | Reading Time: ~12 minutes | Channel: techminute
Close your eyes and imagine your slowest, most frustrating AI wait: the spinner, the ellipsis, the coffee break that isn't. Now imagine that same model — OpenAI's flagship GPT-5.6 Sol, no distilled tricks, no quality cuts — answering at up to 750 output tokens per second. That's roughly 10,000 words a minute. A page of prose while you blink twice.
That's the promise of Ultrafast mode, unveiled in limited preview on August 13, 2026, and it's not a software optimization or a clever quantization trick. It's a different kind of silicon sitting inside a different kind of company — one that bet $10 billion (and maybe twice that) on the idea that the future of AI isn't just smarter models, but models that finally keep up with you.
Here's how a dinner-plate-sized chip, a chip company's dramatic IPO story, and OpenAI's speed-obsessed infrastructure strategy collided to create the most interesting inference launch of the summer.
For years, the frontier-AI game has been a quiet Faustian bargain: the smarter the model, the slower it responds. Reasoning models like GPT-5.6 Sol don't just generate text — they think, allocating internal tokens to plan, verify, and reconsider before producing output. Every thinking token costs time. As models scaled up in size and intelligence, they scaled up in latency too. Users either waited for high-quality answers or accepted cheaper, faster, dumber ones.
OpenAI's own framing in the announcement is blunt: "Until now, getting real-time speed typically meant choosing a smaller or more specialized model." Small models fly but can't reason. Frontier models reason but crawl. The stack was — is — a menu of compromises.
Against that backdrop, OpenAI has spent the past few months turning speed itself into a product tier. The API already offers a "Fast Mode" that runs GPT-5.6 Sol at up to 2.5× speed for roughly double the price, as The Decoder reported. Ultrafast adds a third rung on that ladder — and if speed becomes a competitive bottleneck across industries, tiering lets OpenAI capture the value of every incremental improvement. Think AWS charging more for provisioned IOPS... except the "instance" here is borderline telepathy.
The headline numbers — up to 750 output tokens per second, up to 14× faster than Standard — come from a company that has been unusually contrarian for over a decade: Cerebras Systems.
Where NVIDIA builds small dies and wires thousands of them into clusters, Cerebras does the opposite: it fuses a whole silicon wafer into a single chip. The current third-generation product, the WSE-3, fabricated on TSMC 5nm, spans 46,225 mm² — 57 times larger than NVIDIA's H100. The specs are absurd:
| Spec | Cerebras WSE-3 | NVIDIA H100 |
|---|---|---|
| Chip area | 46,225 mm² | 814 mm² |
| Transistors | 4 trillion | 80 billion |
| AI cores | 900,000 | 16,896 |
| On-chip memory | 44 GB SRAM | 50 MB L2 |
| Memory bandwidth | 21 PB/s | 3.35 TB/s |
| Peak FP16 | 125 PFLOPS | 1.98 PFLOPS |
(Source: Cerebras S-1 coverage; NVIDIA public spec sheets.)
Why it matters for speed: inference is memory-bound — every token requires reading model weights. On clusters, weights live on HBM outside the die and must be shuttled across a bus, a data-movement bottleneck. Cerebras keeps 44 GB of SRAM directly on the wafer; weights stay on-chip and tokens flow through without bus round-trips. The company claims its CS-3 systems serve large models up to 15× faster than comparable GPU solutions — a company claim, not independent.
Cerebras ran its own benchmarking on Humanity's Last Exam, 2,500 PhD-level questions:
That's ~7× wall-clock at comparable accuracy. And GDP-Val, economically valuable tasks (inside Codex): a 5.6× end-to-end speedup with no quality degradation. Artificial Analysis comparisons: Ultrafast is 11× faster than Fable 5 and 5× faster than Opus 4.8 Fast.
| Metric | Claude Fable 5 | GPT-5.6 Sol Ultrafast |
|---|---|---|
| HLE 2,500 questions | 78h 27m | 11h 11m |
| GDP-Val end-to-end | baseline | 5.6× |
| Output tokens/sec | — | up to 750 |
Cerebras self-reported benchmarks; independent replication pending.
At 750 tokens/sec the model doesn't just feel faster — it changes where the bottleneck is. OpenAI researcher Jeffrey Wang: "Whereas formerly I might have to wait a couple minutes for a job to finish, it now finishes before I even get the chance to switch contexts."
Use cases multiplying:
Early users: Jane Street, Podium, Basis, Rogo — sub-second frontier intelligence in production.
Ultrafast sits on a January 2026 partnership: OpenAI committed to add 750 MW of low-latency compute in tranches through 2028. Reuters reported the contract at $10 billion; later (unconfirmed) reports reference $20 billion+ with an equity stake.
Cerebras grew to $510 million revenue in 2025 (~88% YoY), but 86% of that came from two UAE entities (62% MBZUAI, 24% G42). OpenAI is the diversifying force — and the concentration story is not over.
None of it kills the story. It just shouldn't make you declare NVIDIA dead.
Ultrafast isn't the fastest model ever written — it's the smartest model running at a speed where the hardware isn't the bottleneck. That flips the last major trade-off: responsiveness no longer forces a smaller model. If 750 tok/s sustains, the next AI pricing battle isn't about models — it's about who can make the biggest model feel like a thought.
All claims verified against Gold-tier (official announcements) and Silver-tier (trade press) sources; all URLs scraped and confirmed accessible 2026-08-15.
