OutYet reporting
GPT-5.6 Sol gets a Cerebras-backed Ultrafast preview, but access is narrow
OpenAI is testing a much faster serving tier for its flagship GPT-5.6 Sol model. The disclosed throughput is notable, while availability, pricing, and production guarantees remain unsettled.
OpenAI has previewed Ultrafast, an API service tier for GPT-5.6 Sol that it says runs up to 14 times faster than Standard processing and reaches up to 750 output tokens per second through Cerebras. This is a serving change, not a newly detected model release: GPT-5.6 Sol already has its own documented API model ID and alias. The immediate development is lower claimed output latency for the same flagship tier, subject to a limited-preview boundary.
The timeline matters because GPT-5.6 launched as a family with Sol as its flagship tier, Terra as the lower-cost option, and Luna as the faster, lower-cost option. OpenAI's model documentation lists Sol with a 1.05 million-token context window, a 128,000-token maximum output, supported streaming and tool features, and a $5 per million input-token and $30 per million output-token standard price. Ultrafast therefore changes how Sol is served rather than expanding its documented context, tools, or modalities.
The comparison with OpenAI's established Fast mode shows why the preview should be read carefully. Fast mode advertises Sol at up to 2.5 times Standard speed, a 99% latency SLA above 80 tokens per second for eligible Enterprise customers, and short-context prices of $10 input and $60 output per million tokens. Ultrafast's stated ceiling of 750 output tokens per second is much higher, but its announcement gives an up-to figure rather than a percentile latency commitment, and the announcement does not publish a specific Ultrafast price.
For technical teams, the plausible use case is a workflow where streamed output itself is the bottleneck, such as an interactive coding, support, investigation, or research loop. OpenAI describes those categories and says engineers remain responsible for judgment and deployment. The practical limitation is equally important: Ultrafast is available only to a select group while capacity grows, and the announcement does not establish broad access, a production SLA, or a benchmark proving that the speed tier changes model quality. Teams should treat the feature as an early-access serving option, not assume it is generally obtainable or equivalent to a published performance comparison.