⏱️ Lectura: 11 min
OpenAI and Cerebras just announced Ultrafast Mode, the inference layer that turns GPT-5.6 Sol Ultrafast into the fastest frontier model available through an API, reaching up to 750 output tokens per second. The announcement, published on August 13, 2026, marks the first time a frontier lab has offered its flagship model at that speed within its own infrastructure.
📑 En este artículo
Until now, working with large models meant choosing between intelligence and speed. Ultrafast promises to erase that trade-off for tasks where every second counts: production incidents, cyberattack response, or agents that no longer force you to switch tabs while waiting for a reply.
TL;DR
- Cerebras and OpenAI unveiled Ultrafast Mode on August 13, 2026, first available in the OpenAI API.
- GPT-5.6 Sol in Ultrafast mode reaches up to 750 output tokens per second without losing quality.
- It’s 11 times faster than Fable 5 and 5 times faster than Opus 4.8 in Fast mode, according to Artificial Analysis.
- On Humanity’s Last Exam (2,500 questions), Ultrafast answered everything in 11 hours and 11 minutes.
- Fable 5 needed 78 hours and 27 minutes for the same exam, nearly 7 times slower.
- On GDP-Val, Ultrafast achieved a 5.6x end-to-end speedup without losing quality.
- Access is in limited preview, with gradual expansion as capacity grows.
- The speed comes from Cerebras’s Wafer-Scale Engine: 44 GB of SRAM per chip, with no external HBM.
What happened: Cerebras accelerates GPT-5.6 Sol Ultrafast
The announcement came from Cerebras’s official blog, written by Joyce Er, under the title Accelerating GPT-5.6 Sol Ultrafast with OpenAI. It explains that Ultrafast Mode is a new service tier launching first inside the OpenAI API, running on Cerebras hardware. For now, access is limited to a select group of customers, with gradual expansion as installed capacity grows.
The number that sums it all up is 750 output tokens per second. According to data reported by Artificial Analysis, GPT-5.6 Sol in Ultrafast mode runs 11 times faster than Claude Fable 5 and 5 times faster than Claude Opus 4.8 in its Fast mode. The comparison isn’t against just any model: Fable 5 and Opus 4.8 are, today, two of the most widely used reasoning models for complex coding and analysis tasks.
To prove that speed didn’t come at the expense of intelligence, Cerebras ran the Humanity’s Last Exam (HLE) benchmark with GPT-5.6 Sol on Ultrafast. HLE gathers 2,500 doctoral-level questions across chemistry, economics, literature, and other disciplines, designed so that no current model can breeze through all of them. Ultrafast answered all 2,500 questions in 11 hours and 11 minutes using Codex with xhigh reasoning. Claude Fable 5, using Claude Code at the same reasoning level, took 78 hours and 27 minutes, more than three days of continuous compute, to reach comparable accuracy.
Put another way: a job that would take Fable 5 an entire weekend, Ultrafast finishes in a single workday, with equivalent precision and nearly seven times the speed.
Context and history
Cerebras isn’t a new name in this space. The company has spent years betting on an architecture different from traditional GPUs: instead of small chips linked by an external network, it manufactures a single chip the size of an entire silicon wafer (wafer-scale), built from the ground up to move AI data without bottlenecks. It had already been using that approach to offer fast inference for open models like Llama and, more recently, coding models.
What’s new in this announcement is the direct partnership with OpenAI to accelerate a closed frontier model within OpenAI’s own infrastructure. Until now, extreme inference speed (hundreds of tokens per second) tended to be reserved for mid-sized open models, run by specialized providers like Groq or Cerebras Cloud itself. Putting that same speed behind a closed frontier model, with the reasoning quality of GPT-5.6 Sol, is a different kind of leap.
The announcement leans on two quotes within Cerebras’s own post. Rohan Varma, of product at OpenAI, sums it up as a way to keep AI moving at the pace you think, code, and collaborate. Jeffrey Wang, a researcher at OpenAI, says tasks that used to take a couple of minutes now finish before he has a chance to switch tasks, which in practice eliminates the cost of context-switching between parallel agent sessions.
Technical details and performance
The technical reason behind the speed lies in how Cerebras solves the data-movement problem. On a conventional GPU, generating each token requires pulling the model’s weights from HBM memory, off-chip, into the compute chip, over and over, token after token. The bigger the model, the heavier that round trip: the bottleneck isn’t how much the chip can compute, but how long it takes to feed it data.
Cerebras attacks the problem from the chip’s physical design: it packs 44 GB of SRAM directly into each wafer, the fastest memory that exists on a chip, and keeps the model’s weights there without moving them. Tokens flow through the model’s layers, pipelined across wafers, without the constant interruption of fetching weights from outside. This design scales reasonably well as models grow, which according to Cerebras leaves a speed advantage that holds up for future frontier models.
flowchart TD
A["Model weights"] --> B{"Architecture"}
B --> C["GPU: external HBM"]
B --> D["Cerebras: on-wafer SRAM"]
C --> E["Repeated transfer per token"]
D --> F["Weights reside on-chip"]
E --> G["Bandwidth bottleneck"]
F --> H["Tokens flow without interruption"]| Model / mode | Reported speed | Comparison to Ultrafast |
|---|---|---|
| GPT-5.6 Sol (Ultrafast, Cerebras) | up to 750 output tokens/s | baseline |
| Claude Opus 4.8 (Fast mode) | not published by Anthropic | 5x slower, per Artificial Analysis |
| Claude Fable 5 | not published by Anthropic | 11x slower, per Artificial Analysis |
On GDP-Val, a benchmark that measures knowledge-work tasks with real economic value (legal reports, financial models, engineering reports), Ultrafast delivered a 5.6x end-to-end speedup compared to GPT-5.6 Sol in standard mode, without degrading answer quality. Cerebras took that measurement on July 31, 2026, using GPT-5.6 Sol and GPT-5.6 Sol Ultrafast with medium reasoning inside Codex.
💭 Key point: the gain isn’t just raw speed: GDP-Val measures tasks with real economic value, where the historical bottleneck was literally waiting for the model to finish writing.
To confirm whether your account already has access, the simplest check is to list the enabled models:
curl https://api.openai.com/v1/models \
-H "Authorization: Bearer $OPENAI_API_KEY" \
| grep -i "ultrafast"This command lists the models enabled on your account; if you have access to the Ultrafast preview, you’ll see a model identifier with the ultrafast suffix. While access remains limited, the absence of that identifier is the simplest way to confirm your account hasn’t been enabled yet.
How to start testing it
Ultrafast Mode is still a limited preview, so the real first step is requesting access from Cerebras’s page or waiting for OpenAI to enable it on your account. In the meantime, you can get your code ready for the day it activates, using the same SDK you already use for GPT-5.6 Sol in standard mode.
Installing the official OpenAI SDK is the same on Windows, macOS, and Linux, since it runs on pip or npm:
# Windows (PowerShell), macOS, and Linux
pip install --upgrade openai
# alternative with Node.js, same instruction on all three systems
npm install openaiWith the SDK installed, the calling pattern doesn’t change: you keep using responses.create, just pointing to the model identifier OpenAI assigns to the Ultrafast variant once it’s enabled on your account. Here’s a realistic example, measuring tokens per second on the client side:
import time
from openai import OpenAI
client = OpenAI() # reads OPENAI_API_KEY from the environment
start = time.perf_counter()
response = client.responses.create(
model="gpt-5.6-sol-ultrafast", # name subject to change in preview
input="Summarize the latest production incident in 5 bullet points.",
reasoning={"effort": "medium"},
)
duration = time.perf_counter() - start
output_tokens = response.usage.output_tokens
print(f"{output_tokens / duration:.1f} tokens/second")That last print is the most direct way to confirm, on your own account, whether you’re running on Ultrafast or on standard mode: if the number hovers around tens of tokens per second, you’re still on Standard; if it approaches the hundreds, you’re already on Ultrafast.
Impact and analysis
The most cited application in the announcement itself is incident response: teams operating web services can use Ultrafast to find the root cause of an outage and deploy a fix before losing valuable minutes against their SLA. The same applies, with even more urgency, to security teams that need to detect and contain an active attacker before the damage escalates.
But the most interesting change isn’t the speed itself, it’s what it enables: agents that no longer force the user to switch tasks. If an agent takes minutes to respond, the user opens another tab, loses their train of thought, and comes back later to check the result. If the agent responds in seconds, you can work with it without losing focus. That’s exactly what OpenAI’s Jeffrey Wang describes in the announcement: tasks that used to require waiting now finish before there’s time to switch context.
⚠️ Heads up: Ultrafast doesn’t replace standard mode for everything. Cerebras is explicit that Ultrafast should be reserved for work on the critical path, while Standard processing remains the choice for parallelizing common tasks, where individual latency matters less than total cost.
That nuance matters: not every task needs 750 tokens per second. Running hundreds of summaries in batch, indexing documents, or classifying tickets are jobs where total batch throughput matters more than the latency of a single response, and there standard mode remains the most cost-efficient option.
What’s next
For now, Ultrafast Mode is a limited preview: Cerebras says access will keep expanding as the capacity of its installed wafers grows, without yet giving a general availability date. The per-token price for this service tier is also unknown, which makes sense given it’s a compute capacity that’s more expensive to produce than standard inference.
What is clear is the direction: if this wafer-scale architecture keeps its speed advantage as models grow, as Cerebras argues, the question for the rest of the industry (Nvidia, Groq, AMD) stops being just how much intelligence fits on a chip, and becomes how many tokens per second it can sustain without losing it.
Try it yourself: request preview access on the official Cerebras announcement page and keep the snippet above ready for the day it activates on your account.
📖 Summary on Telegram: View summary
Frequently Asked Questions
What is Ultrafast Mode?
It’s a new inference service tier for GPT-5.6 Sol, launched first inside the OpenAI API and running on Cerebras’s Wafer-Scale Engine hardware. It generates up to 750 output tokens per second without sacrificing quality.
How is it different from GPT-5.6 Sol in standard mode?
The model is the same; what changes is the inference infrastructure. Ultrafast runs on Cerebras chips instead of GPUs, prioritizing minimum latency per response over the total cost of processing large batches.
Can I use Ultrafast Mode today?
Only if you’re part of the select group of preview customers. Cerebras and OpenAI plan to expand access gradually as installed capacity grows, with no confirmed general availability date.
What is Humanity’s Last Exam?
It’s an exam of 2,500 doctoral-level questions across disciplines like chemistry, economics, and literature, designed to measure how close AI models are to the edge of verifiable human knowledge.
Does Cerebras compete directly with Nvidia’s GPUs?
Not exactly on the same turf: Nvidia dominates training and general-purpose inference with GPUs, while Cerebras specializes in low-latency inference with its wafer-scale chip, targeting workloads where per-response speed is critical.
What kind of tasks benefit most from Ultrafast?
Ones on the critical path of a problem: diagnosing production outages, responding to security incidents, and agents that need to work in real time without the user having to switch tasks while waiting.
References
- Cerebras: Accelerating GPT-5.6 Sol Ultrafast with OpenAI: the official announcement with the HLE and GDP-Val figures and the Wafer-Scale Engine architecture.
- Artificial Analysis: independent benchmark cited by Cerebras to compare output speed across models.
- OpenAI API Docs: official documentation for the SDK and endpoints used in this article’s examples.
- Wikipedia: Wafer-scale integration: technical background on the single wafer-scale chip manufacturing approach.
📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Omar:. Lopez-Rincon en Unsplash
0 Comments