OpenAI fully opened Decisions API to all customers on October 22 (Thursday). This product debuted at Developer Day on September 29 and now supports both speech and vision modalities, with inference resources reallocated globally to minimize latency
It sounds like just another routine API launch—but the category it occupies deserves separate attentiondecision models— Not generating text, not performing reasoning chains; instead, rapidly presenting a set of options and selecting one, along with confidence scores. Over the past year, this category evolved from TypeSafe’s sole concept into a competitive arena where OpenAI, Microsoft, Amazon, and Liquid AI are all actively competing
OpenAI’s approach: No new model training—just speed up Luna
Most notably, OpenAI’s implementation path: They did not train a new “calibrated decision model,” but instead repurposed ChatGPT Luna Repurpose and reconfigure to accelerate inference speed and reduce per-call cost. Nikunj Handa, OpenAI API Product Lead, told TechCrunch plainly: “We may train specialized models for these capabilities in the future. But right now, it’s the legacy Luna doing the work.”
The primary impact is on latency. Handa cited figures dropping from the original 1–3 seconds to 100, 120, and 180 millisecondsAmong early adopters, Speak’s CTO reported wait time between actions fell from 1.4 seconds to 0.13 seconds. Box, Datadog, and Ramp are also replacing their existing LLM calls with this API to achieve higher throughput.
This approach aligns with the earlier-discussed “routing wizard” logic: many business scenarios don’t require a large language model to generate eloquent text—only to decide “what to do next” within hundreds of milliseconds. Using flagship models for such tasks is like delivering parcels by truck. Microsoft-Decision-1 Real-world rankings: OpenAI is not number one
According to third-party benchmarks, OpenAI entered this market late—and its first iteration holds no advantage:
On two independent evaluation indices,
- TypeSafe’s latest Jev modelranks 7th and 3rd, while OpenAI’s Decision Model ranks8th and 14thEvaluators favor Liquid’s model, released last week, and Base Lab’s model more highly.
- On the “Pac-Man decision-making” benchmark, Jev’s latest version ranks first. D1 Ultra OpenAI’s differentiation lies in Elephant-1
- multimodality. Handa stated Luna is “exceptionally strong at image classification, with superior calibration,” and customers want the API to support images as part of its output. In other words, OpenAI isn’t competing head-on for pure speed rankings; instead, it leverages versatility—“seeing images, hearing audio, and generating visual outputs”—to cover workloads that single-point decision models cannot handle—a strategy distinct from
Jev’s singular-excellence approachmultimodal。Handa 说在图像分类任务上”Luna 极强，校准表现更好”，客户也希望 API 能支持以图像作为输出的一部分。也就是说，OpenAI 不打算在纯速度榜上硬拼，而是用”能看图、能听声音、还能输出视觉结果”的通用性去覆盖那些单一决策模型打不了的工作负载——这一点和 Jev 的单点极致路线形成明显对照。
What this means for developers
几个现实影响：
- 成本结构变了：把每秒都在跑的高频判断（路由、分类、过滤、下一步动作）从旗舰模型迁到决策 API，账单元件从”每千 token”变成”每千次判断”，量级差几十倍
- 交互设计解锁：100 毫秒级判断意味着可以放进实时语音对话、游戏帧循环、Agent 的每个动作节点。Handa 已经在谈 60 毫秒、每帧一次的设想
- 选型没有标准答案：TypeSafe CEO Diogo Almeida 公开表示不信任 benchmark，建议各家在自己的工作负载上实测。校准度、成本、速度三者的最优点因场景而异
- 多模态是真门槛：如果你的判断链路里需要读图（截图分类、UI 元素识别、视觉质检），OpenAI 这条路目前更省事
这个赛道的本质：把”智能”做成水电
TypeSafe 给模型取名 Jev，来自经济学家 Jevons——他提出提高资源使用效率反而会增加总消耗（杰文斯悖论）。这个比喻放在决策模型上很贴切：当一次判断便宜到可以忽略不计，产品就会开始疯狂地判断——每一个动作、每一帧、每一次点击都过一次模型。需求不是被满足，是被创造。
从这个角度看，OpenAI 把 GPT-6.1 Luna 改一改就上线，恰恰说明它判断这波机会在”铺量”而不在”精度”：先用已有模型把开发者的高频调用习惯占住，再靠后续专门训练把质量补上。目前这个赛道的玩家除了 TypeSafe 和 OpenAI，还有开源可本地跑的 Amazon Strands Decider 2B，以及刚拿到 75 亿美元估值的 TypeSafe 资本故事——三条路线（闭源 API、开源本地、专精指标）已经开始分野。
Frequently Asked Questions (FAQ)
OpenAI Decisions API 是什么？
一个面向”快速判断”场景的 API：给模型一组选项，它在 100～180 毫秒内选出结果并给出置信度，替代大模型去做路由、分类、下一步动作这类高频小决策。它基于 ChatGPT Luna 重新配置，不是专门训练的新模型。
决策模型和普通大模型有什么区别？
普通大模型擅长生成文本和做多步推理，但慢且贵；决策模型只做”选哪个”和”有多确定”，速度在百毫秒级，成本低一个量级。适合放进实时交互链路，不适合写文案。
OpenAI 的决策模型排名第一吗？
不是。在第三方决策模型指数上，TypeSafe 的 Jev 排在第 3 和第 7，OpenAI 的 Decision Model 排第 8 和第 14。OpenAI 的优势在多模态（能读图、支持视觉输出）和生态整合，不在单项指标。
普通开发者怎么用上？
Decisions API 已对所有 OpenAI API 客户开放，10 月起补齐了语音和视觉模态。按调用量计费，适合先在路由、意图分类这类低风险环节做对比测试，再决定是否替换现有 LLM 调用。
Summarize
决策模型这条线在半年内走完了”论文概念 → 创业公司 → 巨头跟进 → 价格战”的全流程。OpenAI 这一手不惊艳但很有效：用现成模型 + 基础设施优化，先把延迟打进百毫秒区间，把场景铺开。真正值得盯的是接下来专门训练的版本——以及当判断成本趋近于零时，产品形态会变成什么样。
想继续深挖 AI 模型选型？查看 AI Model Library and Tool Comparison Engine，或阅读：OpenAI DevDay 2026 全盘点 · Liquid AI LFM 2.5 发布 · Replit deprecates Anthropic.
AI Dash — Discover the best AI tools