OpenAI Launches 'Ultrafast' Mode for GPT-5.6 Sol, Hitting 750 Tokens/Second via Cerebras

OpenAI has launched a preview of "Ultrafast" mode, a new inference tier for GPT-5.6 Sol that OpenAI says delivers up to 750 output tokens per second — roughly 14 times the speed of standard inference on the same model.

The acceleration comes from Cerebras, whose specialized inference chips power the new mode under a $10 billion partnership signed with OpenAI earlier this year. Ultrafast is currently available only through the OpenAI API for GPT-5.6 Sol and is limited to select customers, with OpenAI saying access will expand gradually as capacity grows. Companies can sign up for updates through a form.

A third speed tier

Ultrafast is not OpenAI's first attempt to monetize inference speed. The company already offers a "Fast Mode" for GPT-5.6 Sol that promises up to 2.5x faster output at roughly double the standard price. Ultrafast adds a third tier above that, though OpenAI has not disclosed specific pricing for the new mode.

The structure mirrors how cloud providers like AWS charge more for higher-performance compute tiers running the same underlying service. OpenAI is applying the same model to inference: customers pay a premium for latency reduction rather than for a different model or larger context window.

Claimed use cases

According to OpenAI, Ultrafast is built to combine the throughput of smaller models with the reasoning capability of a full-size flagship model, which the company describes as "more useful work per second."

OpenAI points to several scenarios where faster output could matter:

• Incident response: engineers analyzing logs, code changes, and reports in real time during an active outage. OpenAI says it already uses the mode internally for this purpose.
• Finance: evaluating market signals and flagging suspicious transactions as conditions shift.
• Customer support: resolving multi-step inquiries across systems in real time.
• E-commerce: answering product questions and checking inventory before a buyer abandons a purchase.
• Research: turning overnight batch experiments into interactive sessions where researchers can adjust an approach and rerun immediately.

These are OpenAI's own framing of potential applications; none of the performance claims for these specific workloads have been independently benchmarked.

What this means

Ultrafast is not a new model — it's a speed tier layered on top of GPT-5.6 Sol, made possible by dedicated Cerebras hardware rather than a change to the model's weights or capabilities. The move signals that OpenAI increasingly treats inference latency as a billable product dimension, similar to compute tiers in cloud infrastructure.

The strategic logic is straightforward: if faster inference unlocks new real-time use cases — incident response, live fraud detection, interactive research — OpenAI captures a share of the value that speed creates, rather than leaving that margin to third-party inference providers. The Cerebras partnership, reportedly worth $10 billion, gives OpenAI a hardware edge that competitors relying solely on GPU-based inference may struggle to match at the same token throughput. Whether 750 tokens per second holds up under broader, non-cherry-picked workloads once general availability arrives remains to be seen.