Summary: Cerebras and OpenAI launched the GPT-5.6 Sol Ultrafast service tier on August 13, 2026, initially in the OpenAI API on Cerebras silicon and reaching 750 output tokens per second. Its wafer-scale engine keeps weights in 44 GB of on-chip SRAM; reported HLE and GDP-Val runs show 7x and 5.6x speedups with comparable or unslipped quality. Access remains a limited preview with no announced pricing.
Model news usually lands on one note: smarter. What Cerebras and OpenAI put on the table on August 13, 2026 with the GPT-5.6 Sol Ultrafast service tier sounds a different one: faster. The tier arrives first in the OpenAI API, runs on Cerebras silicon, and lifts GPT-5.6 Sol to 750 output tokens per second — with, as the announcement puts it, “no quality compromise.”
Anyone building real-time products already knows that speed is a capability in its own right. Shrink response time from minutes to something close to instantaneous and product shapes that never penciled out suddenly become buildable.
What 750 Tokens per Second Means
Set it beside the output speeds Artificial Analysis tracks: GPT-5.6 Sol Ultrafast is 11x faster than Anthropic’s Claude Fable 5 and 5x faster than Opus 4.8 on Fast mode. At 750 tokens per second, a 3,000-token response is done in roughly four seconds — a latency band that reads as conversation.
Cerebras is deliberate about how it frames this. The story is not a peak-number contest; it is frontier intelligence served at real-time speed for the first time. Capability and inference speed travelled separate roads until now, and those roads are beginning to converge.
Wafer-Scale Engineering: Keep the Weights On-Chip
The architecture is where the speed originates. A wafer-sized chip in Cerebras’ Wafer-Scale Engine carries 44 GB of SRAM, so model weights never have to leave the chip, and tokens move through the layers as a pipeline spanning wafers. Conventional GPU inference stalls on hauling weights into and out of HBM memory, which makes bandwidth the ceiling; remove the hauling and the ceiling goes with it.
Cerebras also claims the design scales gracefully as models grow — a bet on future generations that deserves tracking rather than trust.
The Numbers: 7x on HLE, 5.6x on GDP-Val
Two measurements are worth writing down. On Humanity’s Last Exam, Ultrafast worked through all 2,500 questions in 11 hours 11 minutes; Claude Fable 5 needed 78 hours 27 minutes — about a 7x speedup at comparable accuracy (run through Codex and Claude Code at xhigh reasoning, July 10–15, 2026). And on GDP-Val, the tier posted a 5.6x end-to-end speedup without quality slipping.
“Comparable accuracy” carries the weight here. The speed was not purchased with distillation or a smaller model; it is the same model sitting on hardware where bandwidth stops being the binding constraint.
Real-Time Agents Are the Big Winner
The scenarios Cerebras calls out are sharply chosen: getting to the root cause of a production outage, answering a cyberattack while it unfolds, agent workflows that depend on live interaction, and legal, financial, and engineering work where every minute of response time has a price attached. What they share is an inability to wait — each extra minute of latency shows up as a loss.
Today Ultrafast is still a limited preview, open to a select set of customers and widening as capacity allows, with no pricing announced. The practical step for product planners is to sort out which workflows are latency-sensitive now, so that moving them over is trivial once access arrives.