OpenAI's GPT-5.6 Sol Ultrafast: Speed Claims Explained

OpenAI has introduced Ultrafast mode for GPT-5.6 Sol, a new API service tier built on Cerebras wafer-scale chips, promising dramatically faster inference without sacrificing accuracy. The launch was reported by TechCrunch on August 13, 2026, following benchmarking runs in July.

What's New

Ultrafast is a new mode for GPT-5.6 Sol that runs on Cerebras hardware — wafer-sized chips packing 44 GB of SRAM each. According to OpenAI, the tier is designed to deliver "more useful work per second" rather than forcing developers to trade model size for speed. The company put it directly: "Until now, getting real-time speed typically meant choosing a smaller or more specialized model. Ultrafast points to progress in a new direction."

The mode is currently in preview, available only to a select group of customers, with broader access expected to expand over time.

The Numbers

Both OpenAI and Cerebras cite 750 output tokens per second for GPT-5.6 Sol on Ultrafast. Beyond that baseline figure, the reported performance gains include:

5.6x end-to-end speedup on the GDP-Val benchmark with no quality degradation (per Cerebras)
11x faster than Claude Fable 5
5x faster than Anthropic's Opus 4.8 on its own Fast mode
On Humanity's Last Exam (2,500 questions), GPT-5.6 Sol Ultrafast finished in 11 hours 11 minutes, versus 78 hours 27 minutes for Claude Fable 5 — roughly 7x faster with comparable accuracy

TechCrunch, however, reports Ultrafast can work at 14x the speed of standard processing — a figure notably higher than Cerebras's own 5.6x claim.

Sources differ on the headline speed multiplier. The report does not clarify whether the 5.6x, 11x, 14x, and 7x figures stem from the same benchmark conditions or different test setups, making direct comparison difficult.

Timeline

July 10, 2026: HLE benchmarking for GPT-5.6 Sol Ultrafast
July 13–15, 2026: HLE benchmarking for Claude Fable 5
July 31, 2026: GDP-Val benchmarking
August 13, 2026: TechCrunch reports the Ultrafast launch

Where It Could Be Used

OpenAI points to several practical applications for Ultrafast, including:

Root-causing and resolving production outages for web services
Cyberattack detection and response for security teams
Financial market analysis
Customer service and support
E-commerce

Why Founders Should Care

For startups building latency-sensitive products, Ultrafast could plausibly reduce response times for use cases like security monitoring, incident response, or customer support — areas where speed directly affects user experience or risk exposure. If the comparable-accuracy claims hold up under independent testing, compute-heavy startups may also see cost efficiencies over time, though no pricing has been disclosed yet.

That said, founders should temper expectations. Early access is limited to a select group with unspecified selection criteria, so broader availability timing is uncertain. The gap between the 5.6x and 14x speed claims suggests real-world performance gains could vary meaningfully depending on workload — independent testing before committing product architecture to Ultrafast seems prudent. There's also a structural dependency worth watching: OpenAI's Ultrafast tier relies on a single chip partner, Cerebras, which could introduce supply or infrastructure risk if demand scales faster than that partnership can support.

What's Missing

Several open questions remain unanswered in current reporting: pricing for Ultrafast, the criteria for early access selection, hardware cost and energy consumption, a clear rollout timeline for general availability, and the exact methodology behind the "comparable accuracy" claims versus Claude Fable 5. Founders evaluating Ultrafast for production use should watch for clarity on these points before making commitments.