Source: Cerebras — 2026-08-13
Summary
Cerebras and OpenAI launched a new "Ultrafast" service tier for GPT-5.6 Sol that runs on Cerebras' Wafer-Scale Engine instead of GPUs, hitting 750 output tokens per second — up to 14x faster than the Standard tier. The headline demo: the same model answered all 2,500 questions in the Humanity's Last Exam benchmark in 11 hours 11 minutes on Ultrafast, versus 78 hours 27 minutes for a comparable setup. The speedup comes from the WSE's 44GB of on-chip SRAM, which eliminates the memory-bandwidth bottleneck that normally throttles token generation on GPU clusters.
Key Takeaways
- Ultrafast hits 750 output tokens/sec, up to 14x the Standard GPT-5.6 Sol tier — a service-tier change, not a new model.
- The full Humanity's Last Exam run (2,500 questions) dropped from 78h27m to 11h11m on the same model, just different hardware.
- The speed comes from architecture, not tricks: Cerebras' wafer-scale chip keeps 44GB of SRAM on-die, so token generation isn't waiting on data to move between chip and off-chip memory.
- This is a distribution play as much as a hardware story — OpenAI is now shipping the same model at meaningfully different price/speed points depending on backend silicon.
Reel Script
Hook (18s)
A benchmark that used to take three and a half days now finishes overnight — same AI model, same questions, just different chip underneath it. That's not a smarter model, that's a hardware trick worth understanding.
Core Concept (80s)
Here's the thing everyone glosses over: when a chatbot feels slow, it's usually not because the model is "thinking harder" — it's waiting. Every time a GPU generates the next word, it has to fetch the model's weights from separate memory chips, and that trip back and forth is the actual bottleneck, not the math. Cerebras builds something completely different: a single chip the size of a dinner plate, with 44 gigabytes of memory built directly onto the same silicon as the processing cores. No trip across the board, no waiting. OpenAI took their GPT-5.6 Sol model — unchanged, same weights, same intelligence — and ran it on this wafer-scale chip instead of a normal GPU cluster. That's the entire trick: same brain, faster nervous system.
Hands-On (100s)
The proof is in one benchmark: Humanity's Last Exam, a brutal 2,500-question test designed to be hard even for frontier models. Running GPT-5.6 Sol through the normal Standard tier to answer all 2,500 questions took 78 hours and 27 minutes — over three days. Running the exact same model through the new Ultrafast tier on Cerebras hardware: 11 hours and 11 minutes. That's not a 20% improvement, that's a 7x reduction in wall-clock time for the identical workload. In raw throughput terms, Ultrafast pushes 750 output tokens per second, and Cerebras is claiming up to 14x over Standard depending on the task. If you're building anything that chains multiple LLM calls together — agents, multi-step research, batch processing — the bottleneck stops being "how smart is the model" and starts being "how fast can I get tokens back," and that's a hardware question, not a model question.
Takeaway (25s)
If your product runs long agentic chains or batch LLM jobs, the model you pick matters less than you think — the serving hardware underneath it might be the bigger lever. Worth pricing out an Ultrafast-tier run against your current setup this week and see what a 7x wall-clock cut does to your unit economics.