OpenAI has introduced Ultrafast, a new service tier designed to run its GPT-5.6 Sol model at up to 14 times the speed of standard processing. Powered by AI chipmaker Cerebras, the new mode can generate up to 750 output tokens per second and is aimed at businesses and developers that need advanced AI reasoning with much lower response times.
Ultrafast is launching first through the OpenAI API in a limited preview for a select group of customers. OpenAI says the technology is intended for time-sensitive workloads such as software incident response, financial research, customer support, commerce and live experimentation. The launch marks a significant shift in AI competition, with model providers increasingly trying to improve not only intelligence and accuracy but also how quickly powerful models can produce useful results.
OpenAI Launches GPT-5.6 Sol Ultrafast
Ultrafast is not a completely new AI model.
Instead, it is a new service tier that runs GPT-5.6 Sol using infrastructure designed to dramatically reduce inference latency.
OpenAI says Ultrafast can run GPT-5.6 Sol up to 14 times faster than standard processing.
| Ultrafast at a Glance | Details |
|---|---|
| Model | GPT-5.6 Sol |
| New mode | Ultrafast |
| Maximum output speed | Up to 750 tokens/second |
| Speed improvement | Up to 14X |
| Infrastructure partner | Cerebras |
| Launch platform | OpenAI API |
| Availability | Limited preview |
| Initial users | Select customers |
| Wider rollout | As capacity grows |
| Main focus | Real-time, time-sensitive AI workloads |
The central idea is to reduce the traditional trade-off between model intelligence and response speed.
What Is Ultrafast?
AI models generally face a trade-off between intelligence and speed.
Larger and more capable models require more computing resources, which can increase the time needed to produce an answer.
Developers therefore often have to choose between a powerful model that takes longer to respond and a smaller model that responds quickly.
OpenAI is attempting to reduce that trade-off with Ultrafast.
Traditional AI Choice
Powerful model
↓
Better reasoning
↓
Higher latency
↓
Longer wait
OR
Fast model
↓
Lower latency
↓
Potentially less intelligence
Ultrafast’s Goal
GPT-5.6 Sol intelligence
+
Cerebras infrastructure
↓
Much faster inference
↓
Real-time AI applications
The objective is to make advanced reasoning practical in situations where waiting several seconds or minutes can reduce the value of the result.
GPT-5.6 Sol Can Generate Up to 750 Tokens Per Second
The headline performance figure is 750 output tokens per second.
Tokens are the basic units AI models use to process and generate text.
At that speed, a response that would normally take considerably longer can be generated almost immediately.
The exact experience will vary depending on the workload, response length, reasoning configuration and other technical factors.
Speed Comparison
Standard GPT-5.6 Sol
Normal processing speed
↓
Ultrafast
Up to 14X faster
↓
Peak reported output rate
750 tokens/second
The 14X figure represents OpenAI’s maximum claimed improvement over standard processing, rather than a guarantee that every workload will run exactly 14 times faster.
Cerebras Powers Ultrafast
Cerebras is providing the hardware infrastructure behind Ultrafast.
The company specializes in AI processors designed to accelerate inference and training workloads.
Its approach differs from conventional GPU-based AI infrastructure.
Cerebras uses wafer-scale processors that keep large amounts of model data close to the computing units, reducing some of the data-movement bottlenecks that can slow down inference.
| Cerebras Technology | Role |
|---|---|
| Wafer-Scale Engine | Large-scale AI processing |
| On-chip SRAM | Stores model weights closer to compute |
| High bandwidth | Moves data rapidly |
| Pipelined processing | Keeps tokens flowing through model layers |
| Target workload | Large AI inference |
Cerebras says its current wafer-scale architecture contains 44 GB of SRAM on each wafer-sized chip.
Why Hardware Matters for AI Speed
Running an advanced AI model quickly is not simply a software problem.
Large models contain enormous numbers of parameters, and those parameters have to be accessed repeatedly as the model generates tokens.
Traditional GPU systems can experience bottlenecks when model data needs to move between different layers of memory.
Cerebras is attempting to reduce this movement by keeping more of the model weights on-chip.
Conventional GPU Approach
Model weights
↓
Memory
↓
Compute
↓
Memory movement
↓
Next token
Cerebras Approach
Large on-chip memory
↓
Model weights remain close to compute
↓
Continuous processing
↓
Faster token generation
This architecture is one of the reasons Cerebras can target extremely high inference speeds.
Why 14X Faster AI Matters
Faster AI can change what businesses can realistically do with advanced models.
A model that takes several minutes to analyze a complex task may not be practical for a live customer interaction.
The same model operating at much lower latency could potentially become part of a real-time workflow.
OpenAI is targeting exactly this type of use case.
Speed Changes the Workflow
Slow AI
Question
↓
Wait
↓
Answer
↓
Human reviews
↓
Next action
Fast AI
Question
↓
Immediate analysis
↓
Answer
↓
Action
↓
Next question
↓
Continuous interaction
The difference is particularly important for AI agents, which may need to perform multiple reasoning and tool-use steps before completing a task.
Incident Response Is a Major Use Case
OpenAI says its teams are testing Ultrafast for incident response.
When a major software system fails, engineers need to analyze logs, traces, recent code changes and reports while the incident is still developing.
An AI system can help process this information and identify likely causes.
With slower inference, the usefulness of the system can be limited by waiting time.
Ultrafast is designed to shorten that delay.
AI Incident Response
System alert
↓
Logs collected
↓
GPT-5.6 Sol analyzes evidence
↓
Potential cause identified
↓
Suggested checks
↓
Possible fix prepared
↓
Engineer reviews and deploys
The engineer remains responsible for judgment and deployment, but the AI can accelerate the analysis cycle.
Financial Research Could Become More Real-Time
OpenAI also identifies financial research and security as potential applications.
Financial markets can change rapidly.
Researchers may need to process large quantities of information, compare signals and identify unusual activity while market conditions are still evolving.
A faster reasoning model could potentially analyze information and return results before the underlying situation changes significantly.
Financial AI Workflow
Market data
↓
News and signals
↓
Transactions
↓
AI analysis
↓
Risk assessment
↓
Decision support
The value of faster AI in this environment comes not simply from producing an answer sooner, but from producing useful analysis while the information remains current.
Customer Support Could Become More Interactive
Another potential use is customer support.
Complex customer questions can require an AI system to retrieve information from multiple sources, reason through the issue and determine the appropriate response.
If that process takes too long, customers may experience awkward pauses or delays.
Ultrafast could allow more complicated AI-powered support interactions to happen in real time.
| Customer Support Task | Potential AI Role |
|---|---|
| Understand request | Reason over conversation |
| Find information | Search connected systems |
| Diagnose problem | Analyze multiple factors |
| Recommend solution | Generate response |
| Take action | Use connected tools |
| Continue conversation | Respond with low latency |
This could be especially important for voice-based AI systems.
Commerce Is Another Target
OpenAI also highlights online commerce.
A customer may ask a shopping assistant about a product, inventory, delivery options or alternatives.
The AI may need to perform multiple actions before responding.
Faster inference could allow these interactions to happen while the customer is still actively considering a purchase.
AI Shopping Assistant
Customer asks question
↓
AI checks product information
↓
Checks inventory
↓
Compares alternatives
↓
Personalizes recommendation
↓
Responds immediately
↓
Customer continues shopping
Reducing latency could therefore have a direct commercial benefit.
Ultrafast Is Designed for AI Agents
The launch is closely connected to the growth of AI agents.
A chatbot generally answers one request.
An AI agent can break a goal into multiple steps and use external tools to accomplish it.
For example, a software agent might inspect a codebase, run tests, analyze an error and make changes before returning a result.
Every step adds latency.
If each step becomes substantially faster, the entire workflow can become much more interactive.
Agentic Workflow
User gives goal
↓
Agent plans
↓
Tool call
↓
Model reasons
↓
Tool call
↓
Model reasons again
↓
Action
↓
Verification
↓
Final result
Ultrafast can potentially reduce the time spent at each model-processing stage.
OpenAI Says Early Customers Are Testing It
OpenAI says it is working with an initial group of companies across coding, commerce, financial research, customer support and other interactive applications.
The company is using these early deployments to determine where extremely fast frontier-model inference provides the greatest value.
Early customers cited by OpenAI include Jane Street, Podium, Basis and Rogo.
The companies are testing different applications, including AI assistants, voice systems and financial research.
Jane Street Is Testing Faster AI Workflows
Jane Street is among the companies testing Ultrafast.
The company has highlighted the speed improvement as useful for working with AI models in a more focused and productive way.
For technical and financial organizations that work with large amounts of information, reducing the time between an AI request and a usable result can potentially change how employees interact with AI systems.
Instead of launching a task and waiting, users can potentially work alongside the model in a more continuous process.
Voice AI Could Benefit Significantly
Voice applications are particularly sensitive to latency.
When a person speaks with an AI system, even a short delay can make the interaction feel unnatural.
Longer reasoning tasks can make the problem worse.
Ultrafast could allow more capable models to participate in voice conversations without forcing users to wait for every complex response.
Voice AI
User speaks
↓
AI understands
↓
AI reasons
↓
AI checks information
↓
AI responds
↓
Conversation continues
The faster each stage happens, the closer AI interactions can move toward natural real-time conversation.
Live Research Could Replace Overnight Workflows
OpenAI says its researchers have also been testing Ultrafast for experiments that previously required overnight processing.
A conventional research workflow may involve launching several experiments at the end of the day and reviewing results the next morning.
Faster inference can shorten that loop.
Researchers could potentially run an experiment, review its results, modify the approach and launch another experiment during the same working session.
Traditional Research Cycle
Experiment
↓
Wait hours
↓
Review results
↓
Modify approach
↓
Run again
Ultrafast Research Cycle
Experiment
↓
Rapid result
↓
Review
↓
Modify
↓
Run again
↓
Repeat
This can increase the number of iterations researchers complete during a working day.
Cerebras Reports 5.6X End-to-End Speedup
Cerebras reported that GPT-5.6 Sol Ultrafast delivered a 5.6X end-to-end speedup on its GDP-Val evaluation for economically valuable knowledge-work tasks.
The company said this speed improvement came without a degradation in the measured quality of the results.
Cerebras also conducted a Humanity’s Last Exam comparison involving 2,500 questions.
In its test, GPT-5.6 Sol Ultrafast completed the set in 11 hours and 11 minutes, while Claude Fable 5 took 78 hours and 27 minutes to reach comparable conclusions.
These comparisons were conducted by Cerebras, and performance can vary depending on the models, configurations and workloads being tested.
| Cerebras Test | Result |
|---|---|
| HLE questions | 2,500 |
| GPT-5.6 Sol Ultrafast time | 11h 11m |
| Claude Fable 5 time | 78h 27m |
| Reported comparable-speed advantage | ~7X |
| GDP-Val end-to-end speedup | 5.6X |
The results illustrate the potential impact of faster inference, although benchmark results should not be treated as guarantees of real-world performance.
Speed Does Not Mean a New Intelligence Level
Ultrafast does not represent a completely different intelligence model.
It is GPT-5.6 Sol running through a different service tier and infrastructure configuration.
The major change is the speed at which the model can produce output.
This distinction matters because users should not interpret “14X faster” as meaning the model is 14 times smarter.
What Changes?
Model intelligence
GPT-5.6 Sol
Infrastructure
Cerebras-powered Ultrafast
Output speed
Up to 750 tokens/second
Primary improvement
Latency
The objective is to make the same advanced intelligence more responsive.
Ultrafast Is Not Yet Available to Everyone
The biggest limitation at launch is availability.
Ultrafast is currently in limited preview and is available only to a select group of OpenAI customers.
OpenAI says access will expand as capacity grows.
That means most developers cannot simply switch on Ultrafast today.
The company is using the preview period to learn where the technology delivers the greatest value before expanding access.
OpenAI Is Targeting Enterprise Customers First
The initial focus on businesses is strategic.
Enterprise customers often have workflows where time has a measurable financial value.
Reducing the time required to resolve an outage, analyze financial information or answer a customer could potentially produce direct savings or additional revenue.
Where Every Second Matters
Software
Faster incident response
↓
Finance
Faster analysis
↓
Customer support
More natural conversations
↓
Commerce
Faster purchasing assistance
↓
Research
More experiments per day
These are the types of applications OpenAI is prioritizing during the preview.
The AI Industry Is Entering a Speed Race
For years, AI competition focused heavily on benchmark scores and model intelligence.
That competition is now expanding.
Companies increasingly need models that are:
- More intelligent
- Faster
- Cheaper
- More reliable
- Better at tool use
- Better at long-running agentic tasks
The launch of Ultrafast demonstrates how inference speed itself is becoming a competitive feature.
Cerebras Gains a High-Profile AI Partner
The partnership also represents an important development for Cerebras.
Nvidia currently dominates the AI accelerator market, but companies such as Cerebras are developing alternative hardware architectures.
Providing the infrastructure for one of OpenAI’s most capable models gives Cerebras a high-profile demonstration of its technology.
The partnership could therefore have implications beyond this particular service tier.
Hardware Competition
Nvidia GPUs
↓
Dominant AI infrastructure
+
Alternative architectures
↓
Cerebras wafer-scale systems
↓
Potentially faster specialized inference
The future AI infrastructure market could contain multiple architectures optimized for different workloads.
Faster Inference Could Change AI Economics
The economics of AI depend partly on how much computing is required to generate each response.
If hardware can process tokens more quickly and efficiently, AI providers can potentially serve more workloads using the same amount of time and infrastructure.
For businesses, faster inference can also reduce the cost of delays.
A customer waiting 30 seconds for a complicated answer may abandon an interaction.
A customer receiving the answer almost immediately may continue.
The economic value of speed therefore depends on the application.
Key Numbers at a Glance
14X
Maximum speed increase over standard GPT-5.6 Sol processing
750
Maximum output tokens per second
44 GB
SRAM on each Cerebras wafer-scale chip
5.6X
End-to-end speedup reported on GDP-Val
2,500
Questions in the Humanity’s Last Exam comparison
11h 11m
Time reported for GPT-5.6 Sol Ultrafast to complete the HLE test
78h 27m
Time reported for Claude Fable 5 in the same comparison
4
Early customer examples highlighted by OpenAI: Jane Street, Podium, Basis and Rogo
Limited preview
Current availability status
What Ultrafast Means for Developers
For developers, the most important change is the possibility of building applications around a much more capable model without accepting the latency traditionally associated with frontier models.
This could make real-time coding assistants, voice agents, financial research systems and interactive AI applications more practical.
However, developers will still need to evaluate the cost, availability, reliability and actual latency of their specific workloads.
Maximum token speed does not automatically translate into a 14X improvement in the total time required for every application.
What Ultrafast Means for Businesses
For businesses, the technology could make AI useful in workflows where response time is directly connected to revenue, productivity or risk.
A faster AI agent could potentially analyze an outage before it becomes a prolonged service disruption.
A financial system could process new information while market conditions are changing.
A voice assistant could complete complicated tasks without forcing customers to wait.
The value therefore comes from combining intelligence with speed.
What Ultrafast Means for the AI Market
The launch could push competitors to focus more heavily on inference speed.
If companies can offer increasingly capable models at near-real-time speeds, the distinction between traditional software interfaces and AI interfaces could begin to disappear.
AI agents could become interactive systems that continuously observe information, reason over it and take action.
That would represent a significant change from today’s prompt-and-response model.
The Biggest Challenge Is Scaling
The limited preview indicates that capacity remains a consideration.
Delivering extremely high inference speeds for large models to many customers requires substantial computing infrastructure.
OpenAI will therefore need to expand capacity while maintaining reliability and economics.
The wider rollout will show whether Ultrafast can move from a limited demonstration into a broadly available enterprise service.
What Investors and Developers Will Watch
The next important developments will include:
- Wider customer access
- Pricing and commercial terms
- Real-world latency
- Cost per completed task
- Reliability under heavy workloads
- AI-agent performance
- Voice AI adoption
- Enterprise deployments
- Cerebras infrastructure expansion
- Competition from other fast-inference providers
The most important metric will ultimately be useful work completed per unit of time rather than raw tokens per second.
Looking Ahead
OpenAI’s Ultrafast mode marks a major shift in the AI industry’s focus from intelligence alone toward intelligence delivered at real-time speeds. By running GPT-5.6 Sol at up to 14 times the speed of standard processing and reaching as much as 750 output tokens per second, OpenAI and Cerebras are targeting workflows where latency can determine whether advanced AI is practically useful. The technology is currently available only to a limited group of customers through the OpenAI API, with wider access planned as capacity grows.
The longer-term significance of Ultrafast will depend on whether faster inference can translate into measurable improvements in productivity, revenue and decision-making. If frontier models can reliably operate at near-real-time speeds, AI agents could move deeper into software engineering, financial research, customer service, commerce and other time-sensitive operations. The competition may therefore increasingly be defined not simply by which company has the smartest model, but by which company can deliver the most useful intelligence in the shortest amount of time.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
