Overnight, GPT-5.6 Sol was accelerated 14 times by OpenAI.

Is AI starting to compete on information asymmetry?

In the early hours of August 14, OpenAI, in partnership with AI chip manufacturer Cerebras, officially previewed the brand-new service tier "Ultrafast Mode" for its flagship model GPT-5.6 Sol.

Under this mode, the maximum output speed of GPT-5.6 Sol can reach 750 tokens/s, which is up to 14 times faster than the current inference baseline of about 53 tokens/s in the Standard mode, without any quality degradation. In horizontal comparison, the accelerated GPT-5.6 Sol is 11 times faster than Fable 5, and 5 times faster than Opus 4.8 in Fast mode.

The Ultrafast mode will be launched first in the OpenAI API, and a limited preview version is currently available to some customers.

To this end, OpenAI and Cerebras also made a table comparing the current speed and intelligence of state-of-the-art large AI models, where GPT-5.6 Sol Ultrafast is in an absolute leading position:

In a post on Cerebras' blog, engineers introduced tests conducted on "Humanity's Last Exam (HLE)", where they compared the model running under Ultrafast mode with its direct competitors. As we know, HLE is a challenging model benchmark consisting of 2500 questions that are usually only solvable by doctors in fields such as chemistry, economics, and literature.

GPT-5.6 Sol completed all the questions in Ultrafast mode in only 11 hours and 11 minutes. In contrast, Claude Fable 5 needs 78 hours and 27 minutes, which is more than three days of continuous computing to reach the same conclusion. The GPT Ultrafast mode completes tasks at the cutting edge of human knowledge within a single working day, with similar accuracy and a speed nearly 7 times that of Claude Fable.

As model capabilities continue to improve, the scope of application for fast inference will also expand. GPT-5.6 Sol is OpenAI's best-performing model to date in legal documents, financial modeling, and engineering reports. On GDP-Val, a benchmark for measuring knowledge work tasks of economic value, Ultrafast achieves a 5.6x end-to-end speedup without any compromise in quality, fully demonstrating how faster inference can accelerate the delivery of economically valuable work.

Faster AI processing speeds unlock a host of new possibilities for emerging workflows, and agents can now be deployed on the critical path of problem-solving. OpenAI has listed some application scenarios for you:

• Incident response and reliability: When a critical system fails, AI analyzes application logs, recent code changes, and engineer reports to identify possible causes and help prepare remediation plans while the outage is still ongoing.
• Financial research and security: Analyze market signals, evaluate trades, and identify suspicious activities amid rapidly changing market conditions.
• Customer support and voice: Resolve complex customer issues in real time without interrupting the conversation, even when finding answers requires multiple steps or systems.
• Commerce: Answer product questions, check inventory, provide personalized recommendations, and resolve checkout issues while shoppers are still making up their minds, preventing hesitation from turning into abandoned carts.
• Real-time research and experimentation: Turn research that used to take overnight to complete into interactive working sessions, allowing teams to test ideas, review results, adjust approaches, and run another experiment without breaking the workflow.

Inside OpenAI, developers tested GPT-5.6 Sol in Ultrafast mode, and incident response is one use case of Ultrafast. When an alert triggers, engineers need to build an accurate picture of the incident while systems and evidence are still changing. With Sol-level intelligence, the team can quickly read logs, analyze traces, aggregate conversations, determine next steps to check, and assist in preparing or verifying remediation plans. Ultrafast mode reduces the latency between observing signals, verifying hypotheses, and selecting the next action, while engineers remain responsible for judgment and deployment.

On the research side, the OpenAI team uses Ultrafast to quickly search knowledge bases, query data, and rapidly collect, organize, and aggregate information from different tools. The common process in past research was that team members launched a batch of experiments overnight and checked the results the next morning. With Ultrafast, the discovery process can be shortened to support multiple iterations within a working day.

The breakthrough of the Ultrafast mode lies in breaking the memory bandwidth bottleneck of traditional GPU clusters during the decoding phase of large model inference, and its implementation mainly relies on Cerebras's wafer-scale hardware architecture.

Among AI chip manufacturers, Cerebras's solution is unique: its successive generations of chips are made from entire wafers, integrating a massive number of computing cores and an ultra-high-speed interconnection network on a single piece.

When traditional GPUs run large model autoregressive decoding to generate Tokens, limited by video memory bandwidth, they need to continuously transfer huge model weights back and forth between off-chip HBM and computing cores; at the same time, multi-card partitioning will bring communication latency across chips (PCIe/NVLink).

Each of Cerebras's latest wafer-scale chips (WSE-3) integrates 4 trillion transistors, 125 petaflops of AI computing power, and up to 44 GB of on-chip high-speed SRAM. All model parameters are directly resident in the ultra-high-bandwidth on-chip SRAM, eliminating the waiting time for repeatedly loading weights from off-chip memory.

Of course, as a cutting-edge flagship model, the total parameter size of GPT-5.6 Sol far exceeds the capacity of a single chip. Cerebras also has a "Pipelined across wafers" mechanism, which distributes each network layer of the model on multiple wafers, with the parameters of each layer resident in the SRAM of its respective chip, enabling seamless pipelined transmission of Tokens between wafers.

For a large number of large model users, the 14x speedup of the flagship model means that many tasks that previously had to switch to secondary models (such as Luna and Terra) can now be run directly at full performance with confidence. For Agent tasks that require multiple rounds of tool calls, code generation and debugging, and complex chain-of-thought reasoning, processes that originally took hours can be compressed to finish within minutes.

Faster speed may also mean a shift in the way we use AI:

In the AI community, people are already looking forward to the Ultrafast mode for the Luna and Terra versions, and the only question is whether Cerebras has enough chips to support it.

References:

https://openai.com/index/previewing-ultrafast/

https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultrafast-with-openai

This article is from the WeChat official account "Machine Heart" (ID: almosthuman2014), the author is Machine Heart focusing on large models, and is published with authorization from 36Kr.