## Accelerating GPT-5.6 Sol Ultrafast — What Changed & Why It Matters On **August 13, 2026**, Cerebras and OpenAI jointly announced **Ultrafast Mode**, a new service tier in the OpenAI API that runs OpenAI's flagship model **GPT-5.6 Sol** at up to **750 output tokens per second** — up to **14× faster than Standard processing** — with no quality compromise. It launched as a limited preview for a select group of customers. ### What changed **A new "speed class" for frontier intelligence.** Until now, getting real-time speed meant choosing a smaller or more specialized model. Ultrafast removes that tradeoff — you get the full intelligence of OpenAI's most capable model at interactive speeds. ([OpenAI announcement](https://openai.com/index/previewing-ultrafast/)) **Headline performance numbers** (per Cerebras, benchmarked against Artificial Analysis-reported speeds): - **11× faster** than Claude Fable 5, and **5× faster** than Claude Opus 4.8 on Fast mode. - On **Humanity's Last Exam** (2,500 PhD-level questions), Sol Ultrafast answered all questions in **11h 11m** vs. **78h 27m** for Claude Fable 5 — comparable accuracy nearly **7× faster**. - On **GDP-Val** (economically valuable knowledge work), a **5.6× end-to-end speedup** with no quality loss. ([Cerebras blog](https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultrafast-with-openai), [Cerebras press release](https://investors.cerebras.ai/news-releases/news-release-details/cerebras-powers-ultrafast-mode-openais-gpt-56-sol)) **The technical mechanism — Wafer-Scale Engine.** The speed comes from Cerebras' wafer-scale architecture: **44 GB of SRAM packed onto each wafer-sized chip**, keeping model weights on-chip so tokens flow through pipelined layers without the memory-bandwidth bottleneck that constrains GPU inference (where weights must be shuttled between on-chip and off-chip storage). ([Cerebras blog](https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultrafast-with-openai)) **A deepening partnership.** This is the most visible product of OpenAI's **$10 billion low-latency compute deal** with Cerebras earlier in 2026, and it puts that capacity behind OpenAI's flagship model rather than a smaller one. ([Unite.AI](https://www.unite.ai/cerebras-runs-openais-gpt-5-6-sol-at-750-tokens-per-second-in-new-ultrafast-tier/)) ### Why it matters - **Speed becomes a competitive advantage.** OpenAI frames Ultrafast as enabling "more useful work per second" and turning speed into a strategic edge — the next constraint on AI adoption is how fast frontier models run, not just how capable they are. ([OpenAI announcement](https://openai.com/index/previewing-ultrafast/)) - **New real-time use cases.** OpenAI highlights incident response during live outages, financial research while markets move, real-time customer support/voice, commerce (preventing abandoned carts), and turning overnight research runs into interactive working sessions. ([OpenAI announcement](https://openai.com/index/previewing-ultrafast/)) - **Early enterprise adopters** include **Jane Street, Podium, Basis, and Rogo**, who report the speed changes how they build products (e.g., Podium's voice stack, Basis's synchronous experiences). ([OpenAI announcement](https://openai.com/index/previewing-ultrafast/)) - **Strategic significance for Cerebras.** Landing the serving layer for OpenAI's flagship gives Cerebras a production reference account at the top of the market, validating its contrarian wafer-scale approach against GPU incumbents. ([Unite.AI](https://www.unite.ai/cerebras-runs-openais-gpt-5-6-sol-at-750-tokens-per-second-in-new-ultrafast-tier/)) ### Caveats - The benchmark figures (HLE, GDP-Val, speed comparisons) are **vendor-run evaluations** by Cerebras, not independent results — Unite.AI explicitly flags this. ([Unite.AI](https://www.unite.ai/cerebras-runs-openais-gpt-5-6-sol-at-750-tokens-per-second-in-new-ultrafast-tier/)) - Availability is a **limited preview**; both companies tie broader rollout to capacity growth rather than a fixed date. ([Cerebras blog](https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultrafast-with-openai)) ### Strongest sources - [Cerebras blog — Accelerating GPT-5.6 Sol Ultrafast with OpenAI](https://www.cerebras.ai/blog/accelerating-gpt-5-6-sol-ultrafast-with-openai) - [OpenAI — Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed](https://openai.com/index/previewing-ultrafast/) - [Cerebras press release — Cerebras Powers Ultrafast Mode for OpenAI's GPT-5.6 Sol](https://investors.cerebras.ai/news-releases/news-release-details/cerebras-powers-ultrafast-mode-openais-gpt-56-sol) - [Unite.AI — Cerebras Runs OpenAI's GPT-5.6 Sol at 750 Tokens Per Second](https://www.unite.ai/cerebras-runs-openais-gpt-5-6-sol-at-750-tokens-per-second-in-new-ultrafast-tier/)