GPT-5.6 Sol Ultrafast: 750 Tokens Per Second via Cerebras

GPT-5.6 Sol Ultrafast — OpenAI’s new inference tier for its most intelligent model — generates up to 750 output tokens per second, roughly 14x faster than Standard processing. OpenAI previewed the tier on August 13, powered by Cerebras wafer-scale chips, launching first in the OpenAI API for a select group of customers.

The number changes what “real-time” means for frontier models. Most flagship APIs today stream at tens of tokens per second; 750 puts GPT-5.6 Sol at the speed of a typing assistant rather than a thinking assistant. OpenAI’s framing is explicit: “bringing our most intelligent model to products and workflows where every second matters.”

Key facts

• Ultrafast is a new service tier, not a new model — it runs the existing GPT-5.6 Sol at up to 14x Standard speed, with no quality compromise, per OpenAI and Cerebras.
• Hardware: Cerebras Wafer-Scale Engine (WSE), which keeps model weights on-chip in 44 GB of SRAM instead of streaming them from external memory.
• Availability: limited preview today, API first, for a select group of customers; access expands as capacity grows. No ChatGPT or Codex access announced.
• Pricing: not disclosed. OpenAI has not published a rate card for the tier.
• Positioning: targeted at latency-critical work — incident response, fraud detection, real-time customer support and voice, agentic workflows where a multi-step loop compounds latency.

What changed

OpenAI has spent the last two years pushing model efficiency — the July price cuts on Luna and Terra were the pricing side of that. Ultrafast is the speed side: a frontier flagship running on wafer-scale silicon inside OpenAI’s own API, rather than on GPU clusters.

Cerebras’s architecture eliminates the memory-bandwidth bottleneck that dominates generation latency on GPUs. On the GDP-Val benchmark of economically valuable knowledge work, the company reports a 5.6x end-to-end speedup with no measurable quality drop. OpenAI’s own announcement emphasizes “more useful work per second” — the explicit claim is that developers no longer have to choose between frontier intelligence and interactive speed.

The timing is notable: DeepSeek V4-Pro went GA the same week with its own agent-speed story, and Gemini 3.7 Flash arrived at half the price of its predecessor. The August 13–16 window is the first week in 2026 where every major lab shipped either speed, price, or both.

Why it matters

• The latency ceiling is moving. Multi-step agent loops — plan, call a tool, verify, retry — compound latency at every hop. At 750 tokens per second, a chain that took seconds per step becomes conversational. Cerebras and Google made the same argument for fast open models: when inference is fast enough, you build different products.
• Wafer-scale is a real third option. NVIDIA GPUs and TPUs dominated frontier inference; Cerebras’s WSE — a single chip the size of a silicon wafer — is now running OpenAI’s most capable model, which validates the architecture for frontier workloads, not just small open models.
• Cost direction is unknown, and that matters. With no published rate card, buyers can’t yet price Ultrafast against Standard. OpenAI’s earlier pricing signals — Luna down 80%, Terra down 20% — suggest a premium tier, but the per-token premium is genuinely open.

What to watch

• The rate card. OpenAI says access expands as capacity grows; pricing for the tier is the single most important missing detail.
• General availability timeline and whether ChatGPT or Codex surfaces get Ultrafast — consumer rollout was explicitly not announced.
• Benchmark comparisons at speed. Artificial Analysis and similar trackers will publish cross-vendor tokens-per-second tables; that’s where Ultrafast’s real advantage gets measured against Groq- and Cerebras-hosted open models.
• Cerebras capacity. The 44 GB SRAM-per-chip design has a fixed model-size ceiling; whether the platform scales to larger frontier models will determine if this is a niche or a pattern.

Official source

• OpenAI: Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed (August 13, 2026)
• Cerebras: Accelerating GPT-5.6 Sol Ultrafast with OpenAI (August 13, 2026)

Updated August 16, 2026.