OpenAI Decisions API Now in Public Beta: Typed Outputs, 10x Faster

OpenAI has introduced a new developer interface — the Decisions API — now entering public beta on the GPT-6 Luna model. Compared to the existing Responses API, this new interface focuses on returning structured decision-oriented data, with significant improvements in speed and cost.

What Problem Does the Decisions API Solve?

Traditional LLM APIs (like the Responses API) typically return free-form text, requiring developers to add an extra parsing layer to extract structured information. The Decisions API takes a fundamentally different approach: it returns typed results directly, including probabilities, choices, and scores.

This means that when you need the model to perform classification, scoring, ranking, or selection among a set of options, you no longer need to write elaborate prompt engineering and post-processing logic — the model responds with a predictable data structure. For use cases like automated decision pipelines, content moderation, and recommendation ranking, this design dramatically reduces engineering complexity.

"Typed output" means the data structure returned by the API is explicitly defined before the call is made, with fixed type constraints on every field — for example, a probability value is always a float between 0 and 1, a choice is always one of the enumerated values, and a score always falls within a specified range. The fundamental difference from free-text responses is that developers no longer need regex or secondary LLM calls to extract key information, and they don't have to worry about the model "going off-script" and producing format drift. This design philosophy is consistent with OpenAI's previously released Structured Outputs feature, but the Decisions API goes a step further by defining task semantics at the interface level, rather than merely constraining output format.

What Does a ~10x Speed Improvement Actually Mean?

The key figure from OpenAI is that the Decisions API returns results approximately 10x faster than the Responses API.

The reasoning behind this isn't hard to follow. Generating a complete natural language response requires the model to decode token by token — the longer the output, the higher the latency. Decision-type tasks, by their nature, only need to produce a small structured result: a probability distribution, a set of scores, or a selected option. By eliminating the lengthy text generation step, response latency drops dramatically.

For production environments that require high concurrency and low latency, a 10x speed improvement isn't just a quality-of-life upgrade — it could directly change what system architectures are feasible. Real-time risk control, online recommendations, and decision nodes in interactive Agents are all extremely latency-sensitive.

Pricing Strategy: Pay for Input, Not Output

On the cost side, the Decisions API is priced at $0.10 per million input tokens, with no charge for output.

This pricing structure is worth noting. Conventional LLM APIs typically charge more for output tokens (because generation is more computationally expensive), but the Decisions API waives output fees entirely. Given that the outputs from a decision-oriented interface are inherently small, this strategy is especially friendly to developers with high-frequency call patterns — costs are almost entirely determined by the length of the context you pass in.

In other words, for applications that repeatedly ask the model to make judgments across large sets of candidates, the Decisions API offers a clear economic advantage over traditional generative interfaces.

In mainstream LLM API pricing conventions, output tokens typically cost 3 to 5 times more than input tokens, because the autoregressive decoding process scales linearly with output length and is difficult to parallelize. The reason Decisions API can waive output charges is precisely that its outputs are so small — a probability distribution or score vector might be anywhere from single digits to a few dozen tokens, negligible compared to text responses that can run hundreds of tokens long. This also means the $0.10/million token input price isn't directly comparable to similar generative interfaces; actual per-call total cost still depends on the specific context length used.

Potential Impact on the Developer Ecosystem

From a product positioning standpoint, the Decisions API reflects a broader trend: LLM capabilities are being decomposed into more specialized, more controllable atomic interfaces. In the past, developers had only general-purpose chat or completion endpoints, and had to find their own ways to fit the model into structured tasks. Now, OpenAI has elevated "making decisions" to a first-class concept.

This specialization brings several concrete benefits:

• Greater predictability: Typed outputs reduce the risk of parsing failures and format errors;
• Lower engineering overhead: No need for complex schema constraints and retry logic;
• Optimized for both performance and cost: Latency and pricing are purpose-built for decision-making scenarios.

That said, it remains in public beta and is currently tied to the specific GPT-6 Luna model. Real-world stability, accuracy, and whether it will expand to other models still need to be validated through broader usage.

Summary

The launch of the Decisions API reflects OpenAI's continued investment in productizing and specializing model capabilities. For teams building decision-intensive applications, the combination of faster responses, lower costs, and more structured outputs makes for a compelling proposition. What remains to be seen is whether it can deliver on these promises in real-world production workloads — and whether OpenAI will continue down this path of carving out more specialized APIs from its broader capabilities.