OpenAI Previews Ultrafast Mode Promising 14x Faster GPT-5.6 Sol

Inside OpenAI's Ultrafast Preview for GPT-5.6 Sol

OpenAI has begun previewing a new operating mode called Ultrafast for GPT-5.6 Sol, the company's latest and most powerful model. The company says the mode can process work at up to 14 times the speed of standard GPT-5.6 Sol, delivering as many as 750 output tokens per second. Tokens are the distinct pieces of text a large language model produces as it builds a response.

The pitch is aimed squarely at businesses. OpenAI says the faster mode is intended for corporate workflows such as incident response, customer service and support, financial market analysis, and e-commerce. The underlying message is that speed no longer has to mean switching to a smaller or less capable model.

For now, the preview is limited. OpenAI is making Ultrafast available to a small group of customers first and says access will expand only as capacity grows. The accelerated mode is powered through OpenAI's partnership with chipmaker Cerebras, which positions the launch as much a hardware story as a software one.

Anthropic has offered a fast mode for Claude, but OpenAI is claiming a materially higher speed ceiling. The product is now in preview, so the key test will be whether this performance holds up outside controlled conditions.

What Ultrafast Signals for OpenAI's Enterprise Race

OpenAI is betting speed wins enterprise workloads

By framing Ultrafast around incident response, customer support, financial market analysis and e-commerce, OpenAI is targeting workflows where latency has direct commercial value. A support agent or trading desk tool that can respond 14 times faster may change whether a model is usable in production rather than just in experiments. The reported 750 tokens per second is a throughput claim, not a quality guarantee; the commercial test will be whether GPT-5.6 Sol retains accuracy and reliability at that pace.

Anthropic's Claude fast mode is now the benchmark to beat

OpenAI explicitly compares itself with Anthropic, whose Claude fast mode offers acceleration but not at the speed OpenAI is claiming. That sets up a competitive race on inference speed among frontier model providers. Speed has become a differentiating feature because many enterprise customers are constrained not by raw model intelligence, but by whether outputs arrive quickly enough for live decision-making.

The Cerebras partnership is the hidden enabler

The preview is powered by OpenAI's partnership with chipmaker Cerebras. That matters because the 14x claim depends on specialised hardware and inference infrastructure, not simply a software setting. It also explains why OpenAI is limiting access and linking expansion to capacity growth: chip supply and deployment scale will determine how quickly Ultrafast becomes a broadly available enterprise product.

Who gains and who loses

OpenAI gains if Ultrafast strengthens its enterprise story against Anthropic and other rival model providers. Cerebras gains visibility and a flagship inference workload for its chips. Enterprise customers in the named workflow categories stand to gain if the preview lives up to its claims. The clearest pressure falls on providers whose fast modes or smaller specialised models were previously the default answer for low-latency AI work.

What Enterprise Buyers Should Test in the Ultrafast Preview

For technology and operations teams evaluating the Ultrafast preview, the practical questions are about fit before price.

• If your workload matches OpenAI's named use cases — incident response, customer service and support, financial market analysis, or e-commerce — request early access now, because the preview is limited to a small group and expansion is tied to Cerebras capacity.
• Run side-by-side tests between Ultrafast and standard GPT-5.6 Sol on your highest-volume prompts. Measure completed-task time and error rates, not only tokens per second, before changing production systems.
• Ask OpenAI for pricing and rate limits for Ultrafast. The announcement does not disclose them, but cost per completed task may matter more than raw speed for high-volume support and analysis workflows.
• Do not assume Claude fast mode or smaller specialised models are your only low-latency options; include Ultrafast in your next enterprise model comparison once OpenAI expands access.