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

Introduction

On August 13, 2026, OpenAI introduced Ultrafast, a new inference service tier for GPT-5.6 Sol, the company's flagship model. Ultrafast runs up to 14 times faster than OpenAI's Standard processing tier and generates up to 750 output tokens per second. OpenAI is making Ultrafast available now, but only in a limited preview to a select group of customers through the OpenAI API, with access expanding as capacity grows.

The announcement matters because it reframes what faster AI has historically meant. Rather than trading model quality for speed by shrinking or specializing a model, Ultrafast keeps the full GPT-5.6 Sol model intact and instead relies on specialized inference hardware to extract more throughput from it. OpenAI built Ultrafast with Cerebras Systems, whose Wafer-Scale Engine architecture underpins the speed gains.

Feature Overview

Cerebras Wafer-Scale Engine: Keeping Weights On-Chip

Ultrafast's performance gains trace back to Cerebras' Wafer-Scale Engine (WSE), a chip architecture that stores a model's weights on-chip rather than shuttling them back and forth between on-chip memory and off-chip storage during inference. On conventional GPU-based inference, that memory-to-chip data movement is typically the bottleneck that limits how fast a frontier-scale model can generate tokens, regardless of available compute. By eliminating that bottleneck, Cerebras' hardware lets GPT-5.6 Sol process each inference request with substantially less latency between tokens.

Speed Claims: Up to 14x Faster, Up to 750 Tokens Per Second

OpenAI states that Ultrafast runs up to 14 times faster than its Standard processing tier and can generate up to 750 output tokens per second. Output token throughput is one of the clearest practical measures of how quickly a model can complete a long response, generate code, or work through a multi-step task. A model producing hundreds of tokens per second can finish work that would otherwise take many seconds in a fraction of the time, a difference that matters most in workflows where a person or another system is waiting on the model's output in real time.

GDP-Val Benchmark: Speed Without a Quality Tradeoff

OpenAI also published results from GDP-Val, a benchmark built around economically valuable knowledge-work tasks such as drafting legal briefs, building financial models, and producing engineering reports. On GDP-Val, Ultrafast delivered a 5.6x end-to-end speedup compared to Standard processing, with no loss in output quality. That claim, no quality loss, is central to the announcement: Ultrafast is positioned not as a faster-but-weaker alternative to GPT-5.6 Sol, but as the same model running faster.

Ultrafast vs. Standard Processing: What's Confirmed So Far

Usability Analysis

Ultrafast is not yet broadly available. OpenAI has opened it only to a select group of customers in a limited preview through the OpenAI API, and says access will expand as capacity grows, without committing to a timeline. For customers who do have access, OpenAI points to corporate workflows where response latency directly affects outcomes: incident response, customer service and support, financial market analysis, and e-commerce. In each of these use cases, a model that returns a usable answer in a fraction of the time a Standard-tier request would take can change what's practical to automate in real time versus what still requires a person to wait.

What's missing from the announcement is any detail on how customers outside the initial preview group can request access, how long the preview period will last, or what integration changes are required to route requests to Ultrafast instead of Standard processing. Those operational details will likely determine how quickly Ultrafast moves from a preview capability to something most API customers can actually use.

Pros and Cons

Pros

• Up to 14x faster inference and up to 750 output tokens per second, a substantial throughput gain over Standard processing
• 5.6x GDP-Val speedup reported with no loss in output quality
• Runs the full GPT-5.6 Sol model rather than a smaller or distilled variant
• Clear, latency-sensitive target use cases: incident response, customer support, financial analysis, e-commerce

Cons

• Limited preview only, available to a select group of customers, with no stated timeline for broader availability
• OpenAI has not disclosed Ultrafast-specific pricing, quota limits, region availability, or SLA terms
• Capacity-constrained rollout means most API customers cannot yet test or plan integrations around it

Outlook

If OpenAI expands Ultrafast beyond its current preview, it would mark a shift in how frontier AI providers compete: not only on model capability, but on inference speed as a distinct dimension of the product. Cerebras has built its business around specialized inference hardware, and its role here suggests more AI labs may look to purpose-built silicon, rather than general-purpose GPUs, to differentiate their fastest service tiers. Because Ultrafast preserves GPT-5.6 Sol's GDP-Val quality rather than substituting a smaller model, it also points to a future where a fast mode and a capable mode no longer have to be separate products. How OpenAI prices Ultrafast, and how much capacity it can bring online, will determine whether that approach scales beyond a preview.

Conclusion

Ultrafast is a meaningful infrastructure announcement rather than a new model. It demonstrates that GPT-5.6 Sol can run substantially faster, up to 14x, without sacrificing GDP-Val quality, by relying on Cerebras' wafer-scale hardware instead of conventional GPUs. For now, it's relevant mainly to the select enterprise customers already in the preview group who run latency-sensitive workflows like incident response or customer support. Everyone else should watch for pricing and broader availability details before planning around it.

Editor's Verdict

OpenAI Ultrafast Preview: GPT-5.6 Sol Runs Up to 14x Faster is a workable proposition that fills a clear gap, even if it doesn't fundamentally change the landscape.

The strongest case for paying attention: up to 14x faster inference and up to 750 output tokens per second, a substantial throughput gain over Standard processing. That alone raises the bar for what readers should expect in this space. Reinforcing that, 5.6x GDP-Val speedup reported with no loss in output quality — practical value rather than just headline appeal. The broader signal worth registering is straightforward: Ultrafast keeps GPT-5.6 Sol's full model intact, avoiding the traditional tradeoff of shrinking a model for speed. On the other side of the ledger, one constraint is real rather than a marketing footnote: limited preview only, available to a select group of customers, with no stated timeline for broader availability. It should factor into any serious decision. Layered on top of that, OpenAI has not disclosed Ultrafast-specific pricing, quota limits, region availability, or SLA terms — which narrows the set of teams for whom this is an obvious yes.

For ChatGPT power users, OpenAI API customers, and enterprise teams already running on the OpenAI stack, the smart move is to track its trajectory and revisit once the rough edges are filed down. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.