OpenAI introduces a speed-focused classifier API
OpenAI CEO Sam Altman revealed a new "Decisions API" during the company's Dev Day event on Tuesday. The tool is designed for developers who need fast, inexpensive classification outputs rather than the slower, costlier reasoning typical of large language models.
The API functions by giving OpenAI's Luna model a predefined set of options to choose between. This approach mirrors functionality found in Jev, a model released earlier this month by startup TypeSafe AI. Both tools act as high-speed classifiers, outputting probabilities for specific categories or agent behaviors at a fraction of the cost of standard LLM interactions.
The shift toward "System One" thinking
Altman described the API's value proposition in terms of efficiency and capability retention. "By focusing the model on that choice, we can make it extremely fast while keeping capabilities like image understanding, broad language support, and safety protections," Altman said.
Diogo Almeida, CEO of TypeSafe AI and a former OpenAI engineer, sees the move as validation of their approach. Almeida, who co-invented reinforcement learning techniques, joked on X about the start of "clone wars" but added that OpenAI's interest signals a broader industry shift. He suggested that building in a "System One" compatible way is the future, referencing his company's term for fast, intuitive processing as opposed to deliberate "System 2" reasoning.
The core premise is that standard LLMs are often overkill for routine software tasks. Developers using Jev have reported that augmenting LLMs with these specialized classifiers makes systems faster and cheaper. While OpenAI's Decisions API is currently in limited preview, early industry commentary suggests significant demand for this category of tool.
Calibration and cost efficiency
Not all fast models are created equal. A key challenge lies in how well the outputs correlate with real-world accuracy. Almeida argues that TypeSafe's competitive advantage lies in the synthetic data used to generate statistically useful outputs. "Fast and cheap is very easy, you know," Almeida told TechCrunch last week. "If you want it really fast and cheap, use dice, right? Intelligence is the hard part, and my North Star is always pushing the intelligence-per-dollar Pareto curve."
For professionals in software development and IT infrastructure, this distinction between raw speed and calibrated intelligence is critical. The emergence of multiple vendors offering similar APIs suggests a maturing market for specialized inference layers.
Implications for AI agent security
One immediate application for these high-speed classifiers is monitoring AI agents. Following a series of incidents where autonomous agents misbehaved on the open internet, OpenAI introduced security measures that use separate models to watch for bad actions. However, Altman acknowledged that this approach comes at a "significant compute cost."
Shapor Naghibzadeh, a cybersecurity professional and founder of QueryStory, recently built a demo using Jev to check each agentic action against its assigned task. The system blocks actions with high confidence of being malicious, flags others for review, and permits the rest. Naghibzadeh estimates that monitoring specific agent behaviors costs $2.94 with Jev, compared to $372 when using a frontier LLM.
For those managing AI Agent Courses and automation workflows, the ability to run cheap, high-frequency checks on every action creates a new layer of reliability. This cost differential means organizations could potentially afford to monitor every step an agent takes, rather than sampling or relying on post-hoc audits.
Why this matters for government, insurance, IT, and legal sectors
For IT and development teams, the Decisions API and its competitors offer a path to scale automation without proportional increases in cloud costs. The ability to offload simple classification tasks from expensive LLMs to cheaper, faster models can improve system latency and budget efficiency.
In insurance and legal contexts, where document classification and initial risk triage are common, these tools could streamline workflows that currently require manual review or slower, more expensive AI processing. Government agencies, often constrained by strict budget cycles and data privacy requirements, may find value in models that offer predictable costs and faster response times for routine digital interactions.
Your membership also unlocks:
