A few weeks after rolling out GPT-5.6 broadly, OpenAI is tackling another major challenge in artificial intelligence: latency. With its new Ultrafast mode, the company wants its flagship GPT-5.6 Sol model to generate responses at speeds until now mostly associated with much lighter models. The announced throughput can reach 750 tokens per second, while keeping the same model and therefore, according to OpenAI, without sacrificing its reasoning capabilities.¹ This speedup might seem purely technical, but it actually addresses a key issue for agentic AI. When an agent has to analyze information, call several tools, check a result and then continue its work, every second of latency compounds across the steps. With Ultrafast, OpenAI aims precisely to cut this waiting time and bring AI agents closer to true real-time execution.
750 tokens per second: GPT-5.6 Sol picks up the pace
GPT-5.6 Sol is the most powerful model in the GPT-5.6 family, ahead of Terra and Luna, positioned for complex tasks in programming, research, cybersecurity, science and automation. OpenAI had already announced that it would run on Cerebras infrastructure at speeds of up to 750 tokens per second, far above traditional inference.¹ What makes Ultrafast interesting is mainly that OpenAI does not present this speedup as a new, lighter version of the model. The idea is to run GPT-5.6 Sol on infrastructure specifically optimized for ultra-low-latency inference, preserving its level of intelligence while sharply reducing generation time.
The difference from traditional acceleration modes is significant. Until now, getting a faster response often meant choosing a smaller model or reducing reasoning effort. Ultrafast instead seeks to further decouple intelligence and speed, by optimizing the infrastructure that runs the model. In practice, reaching 750 tokens per second means that long responses can appear almost instantly once generation begins. Actual results, however, depend on many factors, including initial reasoning time, context length, calls to external tools and infrastructure load. OpenAI also notes that real-world performance may vary depending on workloads.¹
Cerebras, the hidden engine behind the speedup
To achieve such performance, OpenAI relies on Cerebras Systems, a US company specializing in processors designed for artificial intelligence. Its approach differs from that of traditional GPU infrastructure: Cerebras builds processors spanning an entire silicon wafer, the Wafer-Scale Engines, to bring memory and compute units closer together and reduce the data transfers that slow down inference.²
This collaboration goes far beyond the launch of a simple fast mode. In 2026, OpenAI and Cerebras signed a multi-year agreement covering the deployment of 750 MW of inference capacity, with the infrastructure ramping up gradually through 2028.² The exact value of the partnership has shifted across announcements and financial filings, but it runs into the tens of billions of dollars, underscoring the strategic importance OpenAI places on fast inference.³ Behind Ultrafast, then, lies a deeper shift in the market: after several years spent training ever more powerful models, labs are now looking to optimize how those models run once deployed.
AI agents could be the first big winners
For a typical conversation with ChatGPT, saving a few seconds mainly adds convenience. For agentic AI, the impact can be much greater. An autonomous agent does not necessarily generate a single response. It may look up information, analyze a file, query a database, run code, check the result and then start again until it reaches its goal. Each step calls on the model and adds latency to the overall workflow.
With much faster inference, these chains of actions can become significantly smoother. A developer could have one agent analyze logs and hunt down an error while another prepares a fix. A financial analyst could tap several sources simultaneously before getting a summary. In customer support or voice interfaces, lower latency could make interactions with artificial intelligence feel much more natural. OpenAI already positions GPT-5.6 Sol as a particularly strong model for agentic tasks and software development, notably thanks to its results on Terminal-Bench, coding agent evaluations and long-running professional workflows.⁴
Speed does not remove every bottleneck, however. An agent still depends on the time needed to access an API, load a document, run a search or obtain human approval. Ultrafast speeds up the reasoning engine, but it does not make the entire surrounding IT system instantaneous. That is why the gains will likely be most visible in applications where inference currently accounts for a large share of total execution time.
A very fast mode, but still far from available to everyone
This is where Ultrafast’s promise runs into its main limitation. Infrastructure capable of running a frontier model at several hundred tokens per second requires considerable hardware capacity. From the launch of GPT-5.6, OpenAI had indicated that access to Sol running on Cerebras would initially be limited to select customers while available capacity was gradually scaled up.¹
This creates a contrast with OpenAI’s broader strategy around GPT-5.6. The family is now much more widely available, with Sol as the flagship model, Terra balancing performance and cost, and Luna as a faster, more affordable option. In the API, GPT-5.6 Sol is priced at $5 per million input tokens and $30 per million output tokens, while Terra costs $2.50 and $15 respectively, and Luna $1 and $6.⁴ Access to certain advanced capabilities also depends on the ChatGPT or Codex plan being used. For example, OpenAI is rolling out GPT-5.6 Sol gradually to eligible plans, while the most advanced reasoning options are reserved for certain subscriptions.⁵
For Ultrafast, the real challenge will therefore be as much economic as technological. If this speedup is expensive, it will mainly make sense for applications where every second has direct value: real-time assistance, interactive software development, incident response, conversational agents, transactions or operational analytics. For background workloads, such as document indexing or overnight analyses, a cheaper standard mode may remain far more sensible.
OpenAI gears up for a new battle over inference
The partnership with Cerebras also reveals a shift in competition among artificial intelligence players. Until now, the race was mainly judged by model size, benchmarks or reasoning ability. Now, the time it takes to obtain that intelligence is itself becoming a competitive advantage.
For OpenAI, working with Cerebras also helps diversify its compute infrastructure in a market historically dominated by Nvidia GPUs. Cerebras claims its specialized architectures can ease certain bottlenecks related to data movement and deliver particularly high inference throughput.² OpenAI is not replacing its existing infrastructure, however, but adding a new category of resources optimized for latency-critical workloads.
This shift could gradually change how companies choose their models. The best AI will no longer necessarily be the one with the highest absolute benchmark score, but the one that offers the best trade-off between quality, cost, token consumption and execution time.
The ethical stakes: when speed also shrinks the time for oversight
Making artificial intelligence ten times faster is not just a technical advance. The faster agents can analyze, decide and act, the more important the question of human oversight becomes. In a typical agentic system, latency still leaves time to check an action, interrupt a process or request approval. When several agents can chain decisions together in a matter of seconds, this window for intervention can shrink considerably.
The issue becomes particularly sensitive in cybersecurity, finance, critical infrastructure and automated administrative processes. GPT-5.6 Sol already has advanced cybersecurity capabilities, and OpenAI acknowledges that evaluations cannot cover every possible combination of models, tools and workflows. The company says it has strengthened its safety mechanisms with real-time controls, monitoring, human and automated red teaming, and risk-based access.¹
A second issue concerns unequal access to computing power. If the fastest versions of the best models remain available only to large companies able to pay for premium infrastructure, a gap could open up between organizations with near-instant agents and those using slower systems. Speed would then become an economic advantage in its own right, just like access to the most powerful models.
Finally, this speedup raises an environmental question. The infrastructure needed to serve frontier models at very high speed requires considerable electrical capacity. The 750 MW agreement between OpenAI and Cerebras illustrates the industrial scale now required to run modern AI systems.² The race for low latency will therefore need to be measured not only in tokens per second, but also in energy efficiency and in the real value produced by each computation.
After intelligence, the race for speed is on
With Ultrafast, OpenAI shows that the next frontier of artificial intelligence is no longer just about making models smarter. It is also about making that intelligence fast enough to keep pace with human activity and IT systems. GPT-5.6 Sol can already reason, code, use tools and coordinate complex workflows. By reaching up to 750 tokens per second on Cerebras infrastructure, these capabilities can now be deployed with dramatically reduced latency.¹
This shift could prove especially decisive for AI agents, whose effectiveness depends as much on their reasoning ability as on the time needed to complete each step. But Ultrafast also highlights the industry’s new tensions: infrastructure costs, unequal access to premium capacity, energy consumption and the need to maintain human oversight as automated systems act ever faster. The next competition between OpenAI, Anthropic, Google and other labs may therefore no longer be only about “which AI is the smartest?” but also about another, increasingly decisive question: how long do you have to wait for it to act?
How does GPT-5.6 Sol’s Ultrafast mode work?
Ultrafast mode is a new execution option developed by OpenAI for GPT-5.6 Sol, designed to sharply reduce latency without swapping the model for a lighter version. Unlike approaches that prioritize speed by reducing reasoning capabilities, Ultrafast keeps GPT-5.6 Sol while optimizing the infrastructure used for its inference. The system can thus reach up to 750 tokens generated per second, dramatically speeding up the production of long responses and the execution of tasks that require multiple successive interactions with the model.
This speedup relies in particular on infrastructure from Cerebras Systems, which specializes in processors dedicated to artificial intelligence. Its Wafer-Scale Engine architecture uses a very large silicon surface combined with memory placed as close as possible to the compute units. This design aims to limit the constant data transfers between the processor and external memory, which are one of the main sources of slowdown during large language model inference.
Ultrafast is particularly significant for agentic AI. An agent does not necessarily produce a single response; it may analyze a request, consult several tools, run code, review a result, fix an error and then continue its work. Each step potentially requires a new call to the model. By speeding up GPT-5.6 Sol, OpenAI therefore aims to shorten these successive loops and bring certain agentic workflows closer to real-time execution.
This approach could be particularly valuable for software development, data analysis, conversational support, financial research and incident response. However, model speed does not eliminate other sources of latency. Calls to external APIs, database access, document loading and human approvals can still slow down the overall workflow. Ultrafast therefore mainly speeds up the part of the process directly tied to GPT-5.6 Sol inference.
- Very high-speed inference: generation of up to 750 tokens per second depending on usage conditions
- GPT-5.6 Sol retained: the flagship model is accelerated with no announced switch to a lighter version
- Cerebras infrastructure: a specialized hardware architecture used to reduce data movement and latency
- AI agent optimization: faster loops for reasoning, tool use, verification and execution
- Software development: the potential to make code generation, analysis and debugging workflows more interactive
- Real-time applications: useful for voice assistants, customer support, operational analytics and incident response
- API integration: a mode designed for professional use and applications leveraging GPT-5.6 Sol through the OpenAI ecosystem
- Ultrafast still depends on large-scale, specialized compute capacity, which may limit its deployment at scale
- Access remains restricted, and wider availability depends on infrastructure capacity
- The announced maximum throughput does not necessarily reflect the speed achieved in every situation
- Calls to tools, APIs, databases or external systems can still slow down agentic workflows
- Faster execution does not automatically guarantee more accurate or reliable responses
- Faster-acting systems require appropriate oversight mechanisms, especially when agents can carry out several successive autonomous actions
- The cost and availability of infrastructure at scale are decisive factors for enterprise adoption of the mode
Learn more
With its Ultrafast mode, GPT-5.6 Sol reflects the search for a new balance between reasoning power, execution speed and a smooth user experience in artificial intelligence models. On a closely related topic, read our article “OpenAI launches GPT-5.5 Instant with fewer errors and faster responses”, which analyzes the previous step in this move toward models that respond faster while improving the reliability of their results.
References
1. OpenAI. (2026). Previewing GPT-5.6 Sol: A Next-Generation Model.
https://openai.com/index/previewing-gpt-5-6-sol/
2. Cerebras Systems. (2026). OpenAI Partners with Cerebras to Bring High-Speed Inference to the Mainstream.
https://www.cerebras.ai/blog/openai-partners-with-cerebras-to-bring-high-speed-inference-to-the-mainstream
3. Reuters. (2026). OpenAI Signs $10 Billion Computing Deal with Nvidia Challenger Cerebras.
https://www.reuters.com/
4. OpenAI. (2026). GPT-5.6: Frontier Intelligence That Scales with Your Ambition.
https://openai.com/index/gpt-5-6/
5. OpenAI Help Center. (2026). GPT-5.6 in ChatGPT.
https://help.openai.com/en/articles/20001354-gpt-56-in-chatgpt/