OpenAI Ultrafast: 750-Token-per-Second Enterprise AI

On August 13, 2026, OpenAI and Cerebras launched a joint preview that redraws the ceiling for enterprise AI inference speed. The new Ultrafast tier runs GPT-5.6 Sol at up to 750 output tokens per second on Cerebras wafer-scale chips, delivering the same frontier intelligence at 14 times the standard processing speed.

For most chat applications, speed is a convenience. For enterprise AI agents running multi-step workflows, latency is a cost that compounds with every iteration.

Why Inference Speed Becomes a Compounding Problem for Enterprise Agents

A single model call carries its latency cost once. A production enterprise agent does not work that way. A standard agentic workflow reads context, calls a tool, evaluates the result, then calls the next tool, repeating the cycle across the full task. Twenty steps is routine in real enterprise deployments. At typical GPU-based inference speeds, that sequence compounds into minute-scale delays that constrain where teams will actually deploy frontier AI.

This is the gap Ultrafast targets directly. As Cerebras put it in its own launch post: “Faster intelligence changes what is possible. With Ultrafast, you can now put agents on the critical path of problems where every second counts.”

OpenAI mapped the Ultrafast use cases to exactly those high-pressure, latency-constrained scenarios: incident response during live outages, real-time voice agents, financial research and fraud detection, and e-commerce support flows. All share one structural property: they run agent loops under live time pressure, where a two-second-per-call overhead previously made frontier-quality models impractical.

How Cerebras Gets to 750 Tokens per Second

The speed advantage traces directly to chip architecture. Standard enterprise AI inference runs on Nvidia GPU clusters, where model weights are loaded from High Bandwidth Memory on every forward pass and shuttled between chips via NVLink or InfiniBand. That data movement is fast by GPU standards, but it is off-chip communication that repeats on every generated token.

Cerebras builds its Wafer Scale Engine 3 (WSE-3) from a single die measuring 46,225 square millimeters, carrying 44GB of on-chip SRAM. The entire model weights, for supported model sizes, reside in that on-chip memory with zero inter-chip communication during decode. According to SemiAnalysis benchmarking, Cerebras CS-3 systems achieve 21 times faster end-to-end latency than Nvidia Blackwell B200 GPUs on single-stream decode workloads.

For GPT-5.6 Sol specifically, OpenAI and Cerebras report 750 tokens per second at full model quality: approximately 560 words per second. At that speed, a voice agent can respond mid-conversation without pause windows, an incident response system can process fresh log data before an outage escalates to the next severity tier, and a 30-second research query returns before an analyst context-switches to something else.

Enterprise Ultrafast Inference: Benchmarks at a Glance

MetricGPT-5.6 Sol StandardGPT-5.6 Sol Ultrafast
Max output speed~50-55 tokens/secUp to 750 tokens/sec
Relative speed1x baselineUp to 14x faster
Underlying hardwareNvidia GPU clusterCerebras WSE-3 (on-chip SRAM)
AvailabilityGenerally availableLimited preview, waitlist
Pricing$5 input / $30 output per M tokensNot yet disclosed
Model qualityFrontier (GPT-5.6 Sol)Same, no degradation
Context window1.05M tokens, 128K outputSame

Who Is Testing Ultrafast First

OpenAI named four early preview customers spanning quantitative finance, customer voice, and financial research software: Jane Street, Podium, Basis, and Rogo.

Courtland Lykins, product lead at Podium, described the change as qualitative rather than incremental. “The speed completely changes the call experience for complex work,” Lykins said. Rogo told OpenAI that analysis which previously felt asynchronous now felt synchronous, the difference between a system users work around and one they interact with directly.

OpenAI is also running Ultrafast internally. Engineering teams use it during live incidents, processing logs and traces before the outage moves to its next phase. Researcher Jeffrey Wang noted that tasks which “used to take a couple of minutes to come back” now return before he can context-switch, keeping work sessions in flow rather than in a queue.

The preview is currently waitlisted, with capacity limits governing access.

The Cerebras-OpenAI Infrastructure Deal

This launch is the first customer-facing product to emerge from a much larger infrastructure commitment. On January 14, 2026, OpenAI announced a 750-megawatt compute agreement with Cerebras, deploying capacity in stages through 2028. The announcement at the time named the 750-token-per-second Sol service as the intended use. The August 13 preview is the activation of that plan.

For Cerebras, the placement is strategically significant. The company went public earlier in 2026 in one of the year’s more prominent listings but has faced market skepticism about whether wafer-scale economics generalize beyond a concentrated set of buyers. Running OpenAI’s flagship API tier is the highest-profile production proof point it has landed. Cerebras shares traded lower on August 13 despite the announcement, a reaction analysts attributed to the preview being previously disclosed in principle rather than a surprise.

For OpenAI, the Ultrafast tier represents a quiet step away from total Nvidia dependence for latency-critical serving, consistent with its broader work on custom silicon and its pattern of diversifying inference infrastructure across GPU clouds and specialized accelerators.

What Ultrafast Changes for Enterprise AI Teams

The core shift is that inference speed is now a separate product axis, not a consequence of choosing a smaller or less capable model. Previously, enterprise teams wanting real-time responses had to accept lower quality. The Cerebras partnership lets OpenAI decouple those variables: the same GPT-5.6 Sol intelligence, at 14 times the output speed.

Voice AI agents can sustain multi-turn conversations at frontier quality without pre-built pause windows or fallback to smaller models. Teams building enterprise voice agents, like those covered in our analysis of enterprise voice AI systems, will find the capability substantially changes what is feasible at production scale.

Incident response automation can process streaming log data and return root-cause analysis within seconds of an alert firing. An Ultrafast-powered agent reading 50,000 log lines during a live outage finishes before standard-speed alternatives would complete even the first parse cycle.

Financial research loops shift from overnight batch to same-session interactive. When a model response arrives before an analyst has context-switched, the research session stays in flow rather than becoming an asynchronous queue to manage later.

Multi-step coding and agentic tasks compress dramatically. A 20-step agent loop at 750 tokens per second completes in roughly the same wall-clock time as two standard-speed calls. This is the structural advantage that matters most for enterprise agentic AI deployment, a topic OpenAI has been building toward since the launch of the GPT-5.6 Sol, Terra, and Luna model tiers earlier this year.

Pricing: The Open Question

OpenAI has not published Ultrafast pricing. Running GPT-5.6 Sol at 14 times the speed on specialist wafer-scale hardware will carry a premium over the standard Sol tier, which sits at $5 per million input tokens and $30 per million output tokens. Early preview customers are operating under non-disclosure agreements on pricing details.

The right framing for enterprise buyers is cost-per-outcome rather than cost-per-token. If Ultrafast resolves a production incident 20 minutes faster than standard speed, preventing SLA penalties, or converts a customer support call that latency-caused frustration would have ended, the economics look different than a raw per-token comparison.

OpenAI’s own positioning reflects this: “More useful work per second” rather than “cheaper tokens.” That framing anticipates buyers measuring Ultrafast against the value of outcomes rather than the price of inference compute.

The Broader Infrastructure Signal

The Ultrafast launch is less about one model tier and more about where AI infrastructure competition is heading. Nvidia still dominates training and general-purpose inference by volume. But Cerebras, Groq, and Together AI have spent three years building the case that latency-optimized silicon wins for specific serving patterns, and a production OpenAI deployment gives Cerebras its clearest proof point yet.

For enterprise teams building agentic systems, the strategic implication is not whether to use Ultrafast today but whether workflows should be designed to take advantage of sub-second frontier inference once capacity opens. Voice agent stacks, incident response pipelines, and financial research systems built around standard latency assumptions will need re-architecting to extract full value from 750-token-per-second responses.

That re-architecting is a strategy question as much as a technical one. Teams at the design stage now will be positioned when Ultrafast, or its successors, reaches general availability. For organizations planning their agentic AI architecture, Enera works with enterprise teams on exactly these latency-sensitive deployment patterns.

Sources: