OpenAI Decisions API Turns AI Judgments Into Yes/No Calls

What happened

OpenAI has launched a new Decisions API designed to classify text and images into simple outputs — yes/no probabilities, category selections, or scale ratings — rather than generating free-form responses. According to The Decoder, the new endpoint runs roughly ten times faster than OpenAI's existing Responses API and is priced at $0.10 per million input tokens.

Alongside the launch, OpenAI has consolidated its paid API pricing structure, reducing the number of tiers from five to three.

Why it matters

The Decisions API reframes what many developers currently use large language models for: not open-ended generation, but structured, repeatable judgment calls — moderation flags, routing decisions, scoring, triage. By stripping the output down to a constrained set of answers and optimising speed and cost accordingly, OpenAI is effectively productising a narrower, cheaper, faster category of AI inference that sits alongside its general-purpose models.

For organisations building AI into operational workflows — content moderation, quality scoring, eligibility checks, support-ticket routing — this signals that classification-style tasks no longer need to be run through more expensive, slower, general-purpose generation endpoints. That has implications for how transformation teams architect AI into high-volume decision points, where latency and unit cost at scale often matter more than linguistic nuance.

By the numbers

• 10x faster processing claimed for the Decisions API compared with OpenAI's Responses API
• $0.10 per million input tokens — the pricing for the new endpoint
• From five to three — the reduction in OpenAI's paid API pricing tiers

The Renascence take

It is tempting to read this as a minor technical SKU update, but the more interesting signal is about where AI decision-making is heading inside real operations — and who gets to govern it.

Most organisations still treat AI output as something a human reviews before it touches a customer. A fast, cheap, binary-decision API quietly removes that friction — and with it, the natural checkpoint where someone asks "should this decision have been made at all?" The behavioral risk isn't the model's accuracy; it's the temptation to automate judgment calls — refunds, access, moderation, eligibility — simply because doing so just got ten times cheaper. Before wiring this into a workflow, service and risk teams should decide which yes/no moments in the customer journey are genuinely low-stakes and repeatable, and which ones still deserve a human in the loop, regardless of what the API now makes technically possible.