On October 22, 2026, OpenAI officially announced that the Decisions API is generally available (GA). This interface is built on GPT‑6 Luna and is explicitly positioned to focus model capability on problems whose answers are "bounded," helping applications and agents make classification, routing and scoring decisions from text, images and structured data at near‑real‑time pace.
According to OpenAI, the Decisions API can respond up to 10× faster than calling GPT‑6 Luna directly via the Responses API. It does not generate long‑form text; instead it returns structured, code‑consumable results—probabilities, selected options or scores—greatly reducing latency and cost.
Core capabilities and design logic
Applications and agents must repeatedly make many small decisions in workflows: how to classify inputs, where to route a request, or what action to take next. These decisions are often simple, but in high‑concurrency, long‑pipeline scenarios latency and cost accumulate quickly.
The Decisions API is designed for precisely these use cases. It runs on a dedicated endpoint (POST /v1/decisions) and currently supports only the gpt‑6‑luna model. Developers provide a shared input (text, images or structured data) and define 1 to 400 questions; the model returns corresponding bounded answers. Supported answer types include:
- Predicate: the probability that a given condition is true;
- Choice: select one option from a preset list, with probabilities for each option;
- Score: provide a rating against an ordered scale.
This design avoids the overhead of having a large model generate free text that must then be parsed, producing directly executable decision results. Since the public preview, OpenAI has added features such as prompt caching, US/EU inference residency, native JSON inputs, embedding images within question instructions, and up to 400 questions per request.
On pricing, OpenAI lists $0.10 per million input tokens, with outputs typically not billed (details such as cache writes and regional premiums still apply), giving it a clear cost advantage for high‑frequency decision workloads.
Direct competition with vertical solutions like Jev
The Decisions API launch is widely seen in the industry as a direct response to specialized decision models such as TypeSafe AI’s Jev. Jev emphasizes low latency, strongly typed outputs and local/edge deployment, aiming for "System 1"‑style quick judgments. OpenAI is entering the same space with a general‑purpose model (Luna) plus a dedicated interface, leveraging multimodal understanding, cloud convenience, compliance options (e.g., HIPAA, Zero Data Retention) and an established developer ecosystem to cover a broader range of scenarios.
Early comparison tests show mixed accuracy results: Jev can be superior in some pure‑text, extremely high‑concurrency scenarios in terms of latency and cost, while OpenAI’s Decisions API has advantages for image inputs, zero‑shot flexibility and cloud‑hosted convenience. The final choice typically depends on the specific workload—whether multimodality is required, whether ultra‑low latency is critical, and how much control over infrastructure is needed.
Practical significance and potential impact
For developers, the Decisions API lowers the barrier to using large models for small decisions. Tasks that previously required complex prompt engineering and post‑processing—classification, routing, prioritization—can now be handled with simpler API calls and much lower latency. It is particularly well suited to intermediate decision nodes in agent workflows, content moderation, customer service routing, and real‑time game/interaction control.
For the industry, this marks a trend of major model providers splitting capabilities more finely: rather than one monolithic model for all tasks, vendors are offering specialized interfaces for high‑frequency, structured, low‑latency needs. This both responds to vertical decision models and fills gaps in their own product portfolios.
Objectively, there are still limits to note. The Decisions API is essentially a focused and accelerated use of Luna’s capabilities, not a separately trained dedicated decision model; in extreme high‑concurrency or ultra‑low‑latency edge scenarios, local or specialized solutions may remain superior. Its performance and cost advantages also depend heavily on OpenAI’s cloud infrastructure and pricing strategy, so long‑term sustainability remains to be seen.
Conclusion
OpenAI’s general availability release of the Decisions API on October 22 is a pragmatic and targeted product move. Rather than pursuing "more general, more powerful," it concentrates GPT‑6 Luna on "bounded" problems, trading breadth for speed and structured outputs that deliver practical application value. For developers building agents and responsive applications, it provides a new, low‑friction decision tool; for vertical players like Jev, competition has moved from existence to differentiation.
As AI capabilities continue to fragment, the decisive factors are often not how smart a model is but whether it can be embedded into real business flows at acceptable cost, latency and interface form. The Decisions API launch is the latest annotation of that trend.
▶️ Enjoyed this article? Creating content is hard!
Subscribe to our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
For more AI news visit:
https://www.airosetta.com