For years, the agentic AI community has faced a frustrating trade-off: the models smart enough to reason through complex tasks in real time were too slow for real time. On August 13, OpenAI unveiled Ultrafast, a new service tier that runs GPT-5.6 Sol up to 14 times faster than Standard processing, delivering as many as 750 output tokens per second. Powered by Cerebras silicon, Ultrafast does not merely make an existing model faster—it fundamentally changes what agentic AI can do in production.
What Ultrafast Mode Actually Delivers
Until now, achieving real-time inference typically meant downsizing to a smaller or more specialized model. That concession is precisely what has held back agentic deployments in time-sensitive domains: incident response, live financial analysis, real-time customer support, and interactive research. Agents operate in tight observe-reason-act loops, and every millisecond of latency compounds across multiple tool calls and reasoning steps. If the underlying model is slow, the agent is slow, regardless of how clever its planning layer is.
Ultrafast removes that constraint. Running on Cerebras hardware, GPT-5.6 Sol can generate output at speeds previously reserved for narrow, task-specific models. According to OpenAI, early tests show the model building a working 3D warehouse simulator from a text prompt side-by-side with the Standard tier—except Ultrafast finishes in a fraction of the time. OpenAI is pitching this not as a niche acceleration feature but as a new default for workflows where “every second matters.”
Why Real-Time Speed Changes the Agentic Equation
The significance of 750 tokens per second becomes clearer when you map it onto how agents actually work. A typical agentic workflow might involve generating a plan, calling an API, parsing the response, reasoning about the next step, generating code, executing it, and summarizing the result. Each of those steps requires model inference. At Standard speeds, a multi-step agent loop can take tens of seconds or minutes. At Ultrafast speeds, that same loop collapses to seconds.
OpenAI has highlighted five early scenarios where this matters:
- Incident response: Analyzing application logs, recent code changes, and engineer reports to identify root causes while an outage is still unfolding.
- Financial research and security: Assessing market signals and suspicious transactions while conditions are still changing.
- Customer support and voice: Resolving complex issues in real time without interrupting the conversation, even when the answer requires multiple tool calls.
- Commerce: Checking inventory, personalizing recommendations, and resolving checkout issues while the shopper is still deciding.
- Live research and experimentation: Turning overnight batch jobs into interactive working sessions where teams can iterate multiple times per day.
In each case, the agent is not running a pre-baked report; it is reasoning live. That difference is what separates agentic AI from traditional automation.
The Industry Is Converging on Agentic Speed
OpenAI is not alone in recognizing that agentic AI needs both brains and velocity. At Google I/O 2026, the company declared the “agentic Gemini era,” rolling out Gemini Spark—a 24/7 personal agent running on dedicated VMs—and Gemini 3.5 Flash, a frontier model optimized for action that outputs tokens four times faster than competing frontier models. Google also introduced Antigravity 2.0, a standalone desktop application for orchestrating cohorts of autonomous agents.
The pattern is unmistakable: the major platform players are simultaneously racing to make their most capable models fast enough to act as real-time agents. Speed is no longer a nice-to-have inference optimization; it is becoming a prerequisite for agentic product design.
Infrastructure Catches Up to the Agentic Era
Meanwhile, the infrastructure layer is adapting to support agentic workloads. On August 11, Databricks announced it is welcoming Electric, the team behind PGlite, a lightweight WASM Postgres database designed to run inside agent sandboxes. PGlite has grown from 1 million to 13 million weekly downloads in twelve months, reflecting surging demand for local, low-latency data primitives that agents can query without round-tripping to a remote database. Combined with Databricks’ Lakebase for centralized state and Electric’s real-time sync engine, the stack gives agent fleets both fast local context and a shared, durable record.
Together, these developments—Ultrafast inference, agent-optimized models, and edge-native data layers—suggest the agentic AI stack is maturing from experiment to production architecture.
What Early Production Use Looks Like
OpenAI has been testing Ultrafast internally with a select group of customers across coding, commerce, financial research, and support. One standout internal use case is incident response. When an alert fires, engineering teams use Ultrafast to read logs, analyze traces, synthesize conversations, and identify or validate fixes—all in a fraction of the time previously required. The model reduces the delay between observing a signal and choosing the next action without removing human judgment from deployment decisions.
For research workflows, OpenAI reports that Ultrafast tightens the experiment-review loop from overnight batch jobs to multiple iterations during the workday. That acceleration mirrors what Google described with Gemini 3.5 Flash and Antigravity: when inference speed and model intelligence scale together, the nature of development work shifts from scheduled to continuous.
Availability and the Road Ahead
GPT-5.6 Sol on Ultrafast mode is available in a limited preview to an initial group of customers, with broader access planned as Cerebras capacity scales. OpenAI is soliciting feedback specifically on where the speed change creates the most value, which suggests pricing and access tiers may evolve based on production usage patterns.
The launch also deepens OpenAI’s partnership with Cerebras, positioning the wafer-scale silicon provider as a critical inference partner for OpenAI’s most demanding workloads. That relationship could reshape competitive dynamics in the inference market, especially if Ultrafast proves sticky for agentic applications where latency is the primary buying criterion.
The Bottom Line
Agentic AI has been bottlenecked by a hard truth: the best models were too slow for real-time work, and the fast models were too limited for autonomous reasoning. Ultrafast mode resolves that tension. By delivering frontier intelligence at 750 tokens per second, OpenAI is not just shipping a faster API tier—it is opening a new category of agentic applications that can operate at the speed of live events. If the preview holds up in production, the speed barrier that has defined agentic AI’s first chapter may finally be gone.
