OpenAI Previews Ultrafast GPT-5.6 Sol API Tier

OpenAI identifies incident response, reliability work, financial research and security, customer support and voice, commerce, and live research and experimentation as candidate use cases. In its announcement, the company argues that prior real-time deployments commonly required smaller or more specialized models, while Ultrafast is intended to provide the capabilities of GPT-5.6 Sol at lower latency.

Sachin Katti, OpenAI's VP of Compute Strategy and GPT-Infra, described the preview in a Cerebras announcement as an effort to learn where reduced latency creates meaningful customer value. "We're starting with a small group of customers to learn where that speed creates meaningful value, and we'll use those learnings to inform how we expand the service over time," Katti said.

Cerebras describes its role as powering the Ultrafast inference service. The company claims that the tier provides the same intelligence as Standard processing, though OpenAI's public announcement primarily specifies the throughput improvement rather than publishing independent quality measurements for the new serving configuration.

Cerebras also published comparative results that should be read as vendor-reported benchmarks. It states that, in its Humanity's Last Exam evaluation, GPT-5.6 Sol Ultrafast answered 2,500 questions in 11 hours and 11 minutes, compared with 78 hours and 27 minutes for Anthropic's Claude Fable 5. Cerebras says the tests used GPT-5.6 Sol Ultrafast with Codex on xhigh reasoning on July 10, and Claude Fable 5 with Claude Code on xhigh reasoning from July 13 to 15.

The company further reports a 5.6x end-to-end speedup on GDP-Val without quality degradation, and cites Artificial Analysis output-speed data to claim a 5x advantage over Claude Opus 4.8 Fast mode and an 11x advantage over Claude Fable 5. Those comparisons are useful indicators of the throughput target, but their applicability depends on model settings, tool use, prompting, output length, concurrency, and the benchmark harness.

For ML teams building interactive systems, output-token throughput is only one part of observed latency. End-to-end response time also includes request routing, prompt ingestion, retrieval, tool calls, safety checks, application-side orchestration, and streaming behavior. Companies deploying comparable high-speed inference tiers often need to measure these stages separately, particularly for voice agents, incident workflows, and multi-step tool-using systems where each model turn can add delay.

The announcement gives Cerebras a high-profile production inference deployment for an OpenAI API tier. Cerebras CEO Andrew Feldman characterized the collaboration as evidence that speed and frontier-model capability need not be mutually exclusive, while OpenAI's materials frame the preview around workloads where delays affect usability.

The limited preview leaves several practical details undisclosed in the published materials, including pricing, regional availability, rate limits, context-window behavior, and formal service-level commitments. Teams considering the tier will need those operating details, alongside workload-specific latency and quality tests, before treating the headline 750-token-per-second figure as an application-level performance guarantee.