OpenAI Ultrafast Mode Hits 750 Tokens/sec on GPT-5.6 Sol
OpenAI announced on 13 August 2026 a preview mode called Ultrafast for GPT-5.6 Sol, its current flagship model, claiming throughput of up to 750 output tokens per second — 14x the speed of standard processing. For developers building latency-sensitive agentic pipelines, the tradeoff between capability and speed has historically meant routing time-critical tasks to lighter models while reserving full GPT-5.6 Sol capability for batch or asynchronous workloads. Ultrafast is OpenAI's explicit argument that this tradeoff is no longer structurally necessary.
OpenAI's framing in its accompanying blog post is pointed: "Until now, getting real-time speed typically meant choosing a smaller or more specialized model. Ultrafast points to progress in a new direction: more useful work per second." At 750 tokens per second, the claimed throughput is high enough to sustain multi-turn agent loops, streaming customer-facing responses, and rapid document summarization without swapping model weights mid-pipeline.
The mode is powered by OpenAI's hardware partnership with Cerebras, a chip company whose wafer-scale architecture is engineered for high-throughput inference. This is not an in-house silicon story — it is a supply-side dependency with direct implications for rollout speed. OpenAI has acknowledged this explicitly, stating that preview availability is limited to a small group of customers and will expand as "capacity grows."
Target Workloads and Competitive Context
OpenAI has flagged incident response, customer service and support, financial market analysis, and e-commerce as primary enterprise verticals for Ultrafast. Each involves either strict wall-clock latency constraints or high request concurrency — conditions where a 14x throughput improvement compounds across a pipeline rather than merely accelerating a single call. For teams already exploring direct corpus interaction patterns for AI agents, moving retrieval-augmented generation loops to a higher-throughput backend is a meaningful architectural change.
Anthropic offers a fast mode for Claude, though OpenAI did not provide a direct token-per-second comparison in its announcement, and the source material supplies no throughput figure for Claude's fast mode.
| Mode | Model | Claimed Throughput | Speed Multiplier | Inference Hardware | Availability |
|---|---|---|---|---|---|
| Ultrafast | GPT-5.6 Sol | Up to 750 tokens/sec | 14x standard | Cerebras (partnership) | Limited preview |
| Fast mode | Claude | Not disclosed | Not disclosed | Not disclosed | Generally available |
Preview Constraints and Capacity Ceiling
The Cerebras dependency is the critical variable governing rollout. Cerebras wafer-scale processors are not commodity supply — production volume is constrained in ways standard GPU clusters are not, meaning OpenAI's ability to onboard enterprise customers at scale hinges on Cerebras manufacturing capacity rather than on OpenAI's own infrastructure decisions alone. Teams evaluating whether to architect new OpenAI API-dependent pipelines around Ultrafast should treat "capacity grows" as an indeterminate timeline and design accordingly — either with fallback routing to standard GPT-5.6 Sol or with throughput targets the current preview tier can actually sustain.
No pricing details for Ultrafast were disclosed. Cost-per-token figures are not available from the announcement.
Frontier-model throughput is becoming a distinct competitive axis, separate from capability benchmarks. OpenAI is betting that 750 tokens per second on a top-tier model changes the deployment calculus for real-time enterprise applications — but the Cerebras supply constraint means the claim and the product's broad availability remain, for now, materially different things.