OpenAI Launches Ultrafast: GPT-5.6 Sol Up to 14x Faster

OpenAI introduced a new speed tier for its API on August 14. The “Ultrafast” mode lets flagship model GPT-5.6 Sol generate up to 750 output tokens per second — roughly 14 times faster than standard operation, according to the company. The acceleration runs on specialized Cerebras hardware, following the $10 billion partnership the two companies struck earlier in 2026.

The move establishes a three-tier system of Standard, Fast, and Ultrafast. Speed becomes its own pricing lever, much like in cloud computing: customers who need real-time responses pay for faster delivery, while batch and background workloads stay cheaper.

In parallel, OpenAI is cutting API prices across the GPT-5.6 lineup. Observers read this as a response to mounting price pressure from Chinese providers offering comparable performance at lower cost — even as DeepSeek simultaneously raises its own prices.

Ultrafast is initially available as a preview through the API and limited to select customers; companies can sign up via an interest form. OpenAI highlights real-time incident response, financial analysis, customer support, e-commerce, and interactive research workflows as target use cases.

For builders, the combination is what matters: falling base prices plus a paid turbo tier. Agentic applications that chain many model calls benefit disproportionately from higher token throughput — exactly where OpenAI is positioning itself against Google and Anthropic.