GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Cerebras Runs GPT-5.6 Sol at 750 Tokens Per Second — 7x Faster Than Fable 5, No Quality Loss

OpenAI and Cerebras are previewing Ultrafast Mode, a new service tier in the OpenAI API. Available initially to a select group of API customers, Ultrafast runs GPT-5.6 Sol on Cerebras’ Wafer-Scale Engine architecture at up to 750 output tokens per second — with no change to benchmark accuracy.

Speed Numbers

ModelSpeedvs Ultrafast
GPT-5.6 Sol Ultrafast750 tok/s
Opus 4.8 Fast~150 tok/s5x slower
Fable 5~68 tok/s11x slower

The gap is structural, not incremental. GPU inference bottlenecks on memory bandwidth: weights must be transferred from HBM to compute units for every forward pass. Cerebras’ WSE eliminates that bottleneck by placing 44GB of on-chip SRAM adjacent to 900,000 AI cores on a single wafer. There is no memory transfer overhead at the scale that limits GPU clusters.

The HLE Test

Cerebras put Ultrafast through Humanity’s Last Exam — 2,500 questions spanning chemistry, economics, and literature at PhD level. GPT-5.6 Sol on Ultrafast completed all 2,500 questions in 11 hours and 11 minutes. Fable 5 on xhigh reasoning needed 78 hours and 27 minutes.

The 7x time advantage held with comparable accuracy. This matters because HLE is exactly the kind of sustained multi-question reasoning workload where queue latency compounds — every token generated feeds subsequent context, so wall-clock time grows with depth of reasoning.

What Ultrafast Changes

Most of the infrastructure built around frontier models assumes tokens take time. Retry logic, timeout budgets, and cost models all encode this assumption. At 750 tok/s, a 10,000-token response arrives in 13 seconds instead of 147 seconds. For agentic workflows where an agent waits on model output before taking the next tool call, that latency reduction compounds across dozens of steps.

Time-sensitive applications — legal review, financial modeling, engineering reports — are the stated targets. GDP-Val, a benchmark testing long-document reasoning, also shows a meaningful improvement per Cerebras.

Constraints

Ultrafast mode is a limited preview. Access is not open. The inference hardware does not yet exist at the scale of GPU clouds — Cerebras has one production-grade wafer fab partnership and Colossus is at 770,000 GPUs with none of those running WSEs at comparable density.

The other constraint is model specificity. WSE silicon is purpose-built per model architecture. Ultrafast currently applies only to GPT-5.6 Sol. Expanding to other frontier models requires bespoke hardware design cycles that GPU clouds do not face.

At 750 tok/s and frontier accuracy, the question is whether the throughput advantage justifies the lock-in. For workflows where latency is the actual constraint — not cost, not quality — the answer is yes.