OpenAI and Cerebras Preview Ultrafast Mode for GPT-5.6 Sol at 750 Tokens/s
OpenAI and chipmaker Cerebras have jointly previewed a new Ultrafast service tier for GPT-5.6 Sol, delivering up to 750 tokens per second—14x faster than standard processing—while retaining the same intelligence, per OpenAI. The tier is in limited preview in the OpenAI API for select customers and will expand as capacity grows, targeting time-sensitive business applications. Cerebras silicon directly powering OpenAI's flagship model API is widely seen as a milestone in the race for faster LLM inference.
Confirmed
- Peak performance: Up to 750 tokens/second, 14x faster than standard processing, with no loss of intelligence.
- How it works: Per QbitAI and Synced (机器之心), the speedup comes from Cerebras' wafer-scale engine, which places model weights directly in on-chip SRAM and bypasses traditional GPU memory-bandwidth bottlenecks.
- Official benchmarks: A financial-terminal-dashboard prompt took 12 minutes 20 seconds in standard mode; the model completed a Humanity's Last Exam run in just 11 hours.
- Competitive comparisons: Per petrusenkomax, the model is nearly 7x faster than Claude Fable 5 on 2,500 PhD-level questions and 5.6x faster on economics knowledge tasks; per eyishazyer, Cerebras claims its chips make OpenAI processing 11x faster than Claude.
- Rollout: Limited preview, first offered in the OpenAI API to a select group of customers, expanding as capacity increases.
- Model optimizations: Per 新智元, OpenAI's GPT-5.6 Builder's Guide shows that switching to the Responses API with "reasoning retention" and "compression" enabled lifts GPT-5.6 Sol's ARC-AGI-3 score to 38.3%.
- Reactions: Peter Diamandis calculated that a 90,000-word novel could be written in under 3 minutes at this speed, arguing the economics of knowledge work will be permanently changed; a reposted thread also quotes a former executive worrying that Hugging Face is in danger.
Unconfirmed
- The specific customer list, detailed pricing, and general-availability timeline remain undisclosed.
Why it matters
- Business value: 750 tokens/second sharply cuts response latency for time-sensitive enterprise use cases.
- Industry trend: Cerebras silicon directly powering the OpenAI API signals deep hardware-software integration and may intensify competition in the LLM inference market; some reposts framed it as AI entering a "speed era."
2026-08-14 ~ 2026-08-15 · 20 related posts
- Episode 1: OpenAI Releases GPT-5.6 System Card: Sol Undergoes 700,000 Hours of Red Teaming(2026-08-13, 2 posts)
- Episode 2: OpenAI and Cerebras Preview Ultrafast Mode for GPT-5.6 Sol at 750 Tokens/s(2026-08-14, 20 posts)
- Episode 3: OpenAI Previews Cerebras-Powered Ultrafast at 750 Tokens/sec(2026-08-16, 3 posts)
Primary sources
- [source] OpenAI Previews Ultrafast Mode: GPT-5.6 Sol Hits 14x Speeds — OpenAI · 2026-08-14
- Cerebras Previews Ultrafast Mode for GPT-5.6 Sol at 750 Tokens/s — kimmonismus · 2026-08-14
- Cerebras Powers OpenAI API Ultrafast Tier, Running GPT-5.6 at 750 Tokens/sec — Sethwinterroth · 2026-08-14
- OpenAI Previews Ultrafast GPT-5.6 Sol; Ex-Staffer Says 'HF is Cooked' — olcan · 2026-08-14
- Cerebras Launches Ultrafast Mode for GPT-5.6 Sol, Boosting Generation Speed by 7x — soumitrashukla9 · 2026-08-14
- OpenAI and Cerebras Preview GPT-5.6 Sol Ultrafast Mode at 750 Tokens/s — 机器之心 · 2026-08-14
- GPT-5.6 Boosts Capabilities & Cuts Costs: ARC-AGI-3 Score Jumps to 38.3% — 新智元 · 2026-08-14
- OpenAI Launches Ultrafast: GPT-5.6 Sol Hits 750 Tokens/sec — 量子位 · 2026-08-14
- GPT-5.6 hits 750 tokens/sec in Ultrafast mode, writes 90k-word novel in under 3 minutes — PeterDiamandis · 2026-08-14
- GPT-5.6 Sol Ultrafast Hits 750 Tokens/s, 7x Faster Than Claude Fable 5 — petrusenko_max · 2026-08-14
- OpenAI's new Ultrafast API runs 14x faster on Cerebras, beating Claude — eyishazyer · 2026-08-15
9 near-duplicate retellings: OpenAI · YeXiu223 · nickbaumann_ · paw_lean · soumitrashukla9 · soumitrashukla9 · threepointone · OpenAI · DeryaTR_