GPT-5.6 Sol Ultrafast API: 750 Tokens per Second on Cerebras
OpenAI's Ultrafast mode runs GPT-5.6 Sol at up to 750 output tokens per second on Cerebras hardware. How it works, who gets it, and what to call today.
On August 13, 2026, Cerebras and OpenAI announced Ultrafast mode: a new OpenAI API service tier that serves GPT-5.6 Sol at up to 750 output tokens per second on Cerebras wafer-scale hardware. OpenAI had been quietly running Sol on Cerebras for select customers since July; the announcement makes the tier official, in limited preview, with pricing still undisclosed. Here is what it is, how the speed happens, and what you can actually call today.
TL;DR: Ultrafast mode is GPT-5.6 Sol served on Cerebras hardware at up to 750 output tokens per second, announced August 13, 2026 as an OpenAI API service tier in limited preview. Same model, same weights, no stated quality trade-off; pricing is not yet public. It is distinct from Sol's ultra mode, which runs four subagents in parallel. Standard GPT-5.6 Sol is callable now as
openai/gpt-5.6-sol, with a 1M-token context and the top Terminal-Bench 2.1 score at 88.8%.
Key takeaways
- Ultrafast mode serves the existing GPT-5.6 Sol model at up to 750 output tokens per second, per Cerebras. It is a serving tier, not a new model.
- The speed comes from Cerebras' Wafer-Scale Engine: 44 GB of on-chip SRAM keeps model weights on the wafer, removing the GPU memory-bandwidth bottleneck.
- Cerebras claims no quality compromise against standard inference; independent verification has not landed yet.
- Access is a limited preview for select customers, expanding as capacity grows. No public API parameter or pricing has been published.
- The name is easy to confuse with Sol's ultra mode (four parallel subagents). They are unrelated controls, and this post keeps them apart.
What is Ultrafast mode?
Ultrafast mode is the same GPT-5.6 Sol that tops the Terminal-Bench 2.1 board, served from different hardware. Nothing about the model changes: same weights, same 1M-token context, same reasoning efforts. What changes is where inference runs. Instead of GPU clusters, requests route to Cerebras wafer-scale systems that generate output several times faster than standard serving.
Cerebras frames the gain in developer terms: an agent loop that streams 20,000 tokens of code finishes in under half a minute instead of several. OpenAI researcher Jeffrey Wang put it more plainly: "It now finishes before I even have the opportunity to context-switch."
How 750 tokens per second works
The bottleneck in normal LLM serving is memory bandwidth: every generated token requires streaming model weights from GPU memory. Cerebras' Wafer-Scale Engine sidesteps this by keeping weights in 44 GB of SRAM on the wafer itself, so tokens flow through pipelined model layers without leaving the chip. For a model Sol's size the weights span multiple wafers, pipelined together.
The claimed results, and these are Cerebras' own comparisons: 11x faster than Claude Fable 5, 5x faster than Claude Opus 4.8 in Fast mode, and a benchmark stunt worth repeating, running the 2,500-question Humanity's Last Exam suite in 11 hours 11 minutes against 78+ hours for Fable 5. Treat the ratios as vendor numbers until third parties publish measurements; the 750 tokens-per-second headline is the concrete claim to hold them to.
Ultrafast mode is not ultra mode
OpenAI now has two similarly named Sol controls, and they do different jobs:
| Ultrafast mode | Ultra mode | |
|---|---|---|
| What it is | Serving tier on Cerebras hardware | Four subagents working in parallel |
| What it changes | Output speed, up to 750 tok/s | Task strategy and quality |
| Token spend | Same tokens, delivered faster | Several times a standard call |
| Benchmark effect | None claimed on scores | Terminal-Bench 2.1: 88.8% to 91.9% |
| Availability | Limited preview | Live on standard serving |
They compose in principle: ultra mode's parallel subagents each stream faster on Ultrafast serving. But today only one of the two is generally available.
How to get access
Ultrafast mode is in limited preview for select OpenAI customers, expanding "as capacity grows." No public waitlist mechanics, API parameter name, or pricing have been published. If your product depends on maximum tokens per second this quarter, the honest options are asking OpenAI directly or building against the fastest generally available serving and swapping later.
What you can call today
Standard GPT-5.6 Sol is live on Unifically as openai/gpt-5.6-sol: 1,050,000 tokens of context, 128k max output, text and image input, and the scores that made it the agent benchmark to beat, 88.8% on Terminal-Bench 2.1 and 80 on the Artificial Analysis Coding Agent Index. Ultra mode and the max reasoning effort work through the standard request shape.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "openai/gpt-5.6-sol",
"messages": [
{ "role": "user", "content": "Audit this repo for flaky retry logic and propose a fix plan." }
],
"stream": true
}'
Billing is pay-per-use per token with no subscription; live rates are on the pricing page, and new accounts start with $0.20 of free balance. Full parameter docs: docs.unifically.com/models/llm/openai/gpt-5.6-sol. If raw speed matters more than Sol-class depth, Gemini 3.7 Flash streams around 340 tokens per second and is generally available now.
Frequently asked questions
What is GPT-5.6 Sol Ultrafast mode?
An OpenAI API service tier announced August 13, 2026 that serves the existing GPT-5.6 Sol model on Cerebras wafer-scale hardware at up to 750 output tokens per second. Same model and weights; only the serving speed changes.
Is Ultrafast mode available now?
Only in limited preview for select OpenAI customers. There is no public API parameter, waitlist, or pricing yet; Cerebras says access expands as capacity grows.
Is Ultrafast mode lower quality?
Cerebras says no, claiming identical output quality to standard inference since the model and weights are unchanged. Independent measurements have not been published yet.
What is the difference between Ultrafast mode and ultra mode?
Ultrafast mode is a hardware serving tier that makes tokens arrive faster. Ultra mode is a Sol capability that coordinates four subagents in parallel, lifting Terminal-Bench 2.1 from 88.8% to 91.9% while multiplying token spend. Unrelated controls with confusable names.
Can I use GPT-5.6 Sol without Ultrafast access?
Yes. Standard GPT-5.6 Sol is generally available, including on Unifically as openai/gpt-5.6-sol with pay-per-use billing and the full 1M-token context window.
What to watch next
Three signals worth tracking: the pricing announcement, since a speed tier is only interesting once its premium is known; the API surface, when OpenAI documents how requests opt into the tier; and whether Ultrafast serving reaches OpenAI's resale and partner channels once it leaves preview. We will update this post at each step, and the GPT-5.6 Sol page will note it if Ultrafast serving ever reaches the platform.
