Google and OpenAI Pivot to Speed With Ultra-Fast AI Model Tiers
The industry is shifting focus from raw intelligence to latency as real-time autonomous agents become the next frontier.
Google and OpenAI have unveiled new AI models and API tiers optimized for extreme speed, signaling a strategic pivot toward real-time agentic workflows. This move marks a transition in the AI race: the primary bottleneck is no longer just the quality of reasoning, but the speed of the iteration loop.
Google has released Gemini 3.7 Flash, a model specifically engineered to support coding and autonomous agent workflows. Simultaneously, OpenAI has previewed "Ultrafast," a new API tier for its GPT-5.6 Sol model. Technical specifications indicate the Ultrafast tier can reach approximately 750 tokens per second, representing a speed increase of up to 14x over standard processing. While Gemini 3.7 Flash is already widely available to developers via Gemini Spark, OpenAI's Ultrafast mode remains restricted to an invite-only preview phase.
The Shift Toward Agentic Latency
For several years, competition between leading AI labs centered almost exclusively on "intelligence"—the ability of a model to solve increasingly complex reasoning tasks. However, the industry is now moving toward the deployment of AI agents capable of performing multi-step tasks autonomously. For these agents to operate effectively in real-time, the latency between a model's thought and its action must be minimized. When an agent is coding or debugging software, any human-perceivable lag breaks the utility of the tool, making throughput a critical performance metric.
Why Speed is the New Feature
This shift suggests that speed is becoming a primary product feature rather than a secondary optimization. High-throughput models enable a new class of applications: autonomous agents that can operate software, write code, and iterate on errors without the delays typical of larger, slower models.
Furthermore, the technical implementation of these speeds reveals a deeper trend in infrastructure. OpenAI's Ultrafast tier is powered by Cerebras hardware, indicating that software-level optimization has reached a plateau. This partnership suggests that specialized hardware is now essential to achieve the throughput required for the next generation of AI productivity tools, moving the battleground from algorithmic efficiency to hardware acceleration.
What to Watch
As these models move from preview to general availability, the industry will monitor whether this speed comes at the cost of reasoning capabilities. While Gemini 3.7 Flash is already in the hands of developers, the broader impact of GPT-5.6 Sol Ultrafast will remain unclear until OpenAI expands access beyond its invite-only group. The coming months will likely determine if specialized hardware partnerships, like the one between OpenAI and Cerebras, become the standard blueprint for all major AI labs seeking to power real-time agents.