The performance numbers put Ultrafast in a class of its own compared to conventional GPU-based deployments:
Cerebras also published a run of Humanity's Last Exam, a 2,500-question benchmark spanning graduate-level subjects. GPT-5.6 Sol Ultrafast completed the full question set in 11 hours and 11 minutes; Claude Fable 5 needed 78 hours and 27 minutes to reach comparable conclusions .
Cerebras's WSE-3 (Wafer-Scale Engine 3) is the key enabler. Unlike conventional multi-GPU clusters that stitch together dozens or hundreds of small chips, the WSE-3 is a single, massive silicon wafer the size of a dinner plate. This eliminates the inter-chip communication bottlenecks that plague distributed GPU systems .
The architecture includes 44 GB of on-chip SRAM, which allows the entire GPT-5.6 Sol model to fit within the processor's local memory . During inference, model weights never need to travel to slow off-chip DRAM, removing the memory-bandwidth wall that typically throttles large language model decoding
.
Concretely, this means a 3,000-token agent turn that takes about 43 seconds on a standard H100 cluster completes in roughly 4 seconds on Cerebras hardware .
Despite the landmark technical announcement, Cerebras stock (NASDAQ: CBRS) fell sharply on the day. The share price dropped 11.85% in regular trading to close at $231.01, and fell as much as 17% in after-hours trading .
However, the decline was largely attributed to Cerebras's Q2 2026 earnings report released the prior evening (Wednesday, August 12) — record revenue of $209.9M (more than double YoY) but a much wider-than-expected adjusted loss of -$2.98 per share, compared with analysts' estimate of -$0.18 .
Analyst reactions were mixed. Morgan Stanley raised its target price on CBRS from $273 to $279, while Citi maintained a Buy rating but cut its target to $320 .