The short answer
Microsoft-Decision-1 and the OpenAI Decisions API do the same job. Each is a hosted decision model: you send content, a question and a fixed set of answers, and you get back a probability for each answer instead of generated text.
Microsoft-Decision-1 costs $0.042 per million input tokens, under half of OpenAI’s $0.10, and Microsoft says it is faster and more accurate. Microsoft ran every one of those comparisons itself. The OpenAI Decisions API reads images, has a far larger context window, a documented request format and official SDKs. Microsoft has not published Microsoft-Decision-1’s request format or endpoint as of 10 October 2026, so you cannot yet write code against Foundry from the docs alone.
Microsoft-Decision-1 vs OpenAI Decisions API at a glance
| Microsoft-Decision-1 | OpenAI Decisions API | |
|---|---|---|
| Maker | Microsoft | OpenAI |
| Model underneath | Post-trained Qwen3.5-9B, 5B to 15B parameters | GPT-6 Luna (gpt-6-luna), size not published |
| Weights | Closed, hosted only | Closed, hosted only |
| Status | Generally available since 9 October 2026 | Public beta since 6 October 2026 |
| Where you call it | Microsoft Foundry; OpenRouter as microsoft/microsoft-decision-1 |
POST https://api.openai.com/v1/decisions; OpenRouter as openai/gpt-6-luna-decisions |
| Request format | Not published for Foundry as of 10 October 2026 | Documented: input plus a questions array |
| Question types | Yes/no, multiple choice, rating | predicate, choice, score |
| Input | Text only | Text, and images as base64 data URLs |
| Context | 32,768 tokens | 1,050,000 tokens on OpenRouter; OpenAI’s docs give no figure |
| Input price per million tokens | $0.042, output free | $0.10, output free |
| Calibration | Microsoft says the probabilities are calibrated | OpenAI makes no calibration claim |
| Median latency | 85 ms, Microsoft’s own run through Foundry | 298 ms on JevBench, measured 6 October 2026 |
| SDKs | None named by Microsoft | OpenAI’s official SDKs, including Python and JavaScript |
Sources: Microsoft’s launch post and Foundry model card, OpenAI’s Decisions guide and changelog, OpenRouter’s models and endpoints API, and Benchmark Heaven’s JevBench data, all checked 10 October 2026.
Is “GPT-6 Luna Decisions” the OpenAI Decisions API?
Yes. Microsoft’s charts compare its model with a row named GPT-6 Luna Decisions, and the appendix of the launch post links that name to OpenAI’s Decisions guide. OpenAI’s guide does not use that name. It calls the product the Decisions API and says gpt-6-luna is the only model it accepts. GPT-6 Luna Decisions is the name OpenRouter gives the same product: OpenRouter lists openai/gpt-6-luna-decisions as GPT-6 Luna served through OpenAI’s Decisions API, by OpenAI, at OpenAI’s price.
The latency Microsoft shows for it, 300 ms, matches JevBench’s row for “OpenAI Decisions (gpt-6-luna)”, which Benchmark Heaven ran against POST /v1/decisions on 6 October. So the chart row, the OpenRouter listing and OpenAI’s beta are the same product.
One other OpenAI model appears in the launch post and is a different thing. Microsoft’s Copilot team found Microsoft-Decision-1 competitive with “GPT5.6 Luna” and 100 times faster. That is an older text model, not the Decisions API.
What each one returns
Both return probabilities over a fixed set of answers and write no prose. Microsoft lists yes/no, multiple-choice and rating questions, plus grading an AI response or agent action against a rubric. OpenAI documents three question types: a predicate returns the probability that a condition is true, a choice returns one of your values with a probability for each, and a score returns a probability-weighted average over ordered levels. On this site those are Noul, Choice and Score.
OpenAI’s request is documented. You send input as text or as user messages with images, and a questions array where each item has a name. Answers come back in an answers array, and a question can come back as a refusal. Microsoft has not published the Foundry request format, the endpoint path, or limits on questions, options or rating levels.
OpenRouter gives a partial way round that. Its alpha Decisions endpoint takes one request shape, with model, state and a keyed map of questions, and it lists both models with decision output. OpenRouter’s Python, TypeScript and Go SDKs have a method for it. This site has not tested either model through that endpoint.
What Microsoft’s benchmarks show
These are vendor-run. Microsoft picked the benchmarks, ran the models and published the charts. Nobody else has reproduced them as of 10 October 2026.
| Microsoft’s run | Microsoft-Decision-1 | GPT-6 Luna Decisions |
|---|---|---|
| Average accuracy, 36 benchmarks, 147,137 questions | 83.5% | 79.4% |
| Calibration score, 100 is perfect | 92.2 | 89.9 |
| Median latency per request | 85 ms, Microsoft’s own measurement | 300 ms, taken from JevBench v1.6.1 |
Microsoft-Decision-1 is first on accuracy in Microsoft’s chart. Quyet-1.0-Large is second at 81.9% and has the best calibration score, 93.1. GPT-6 Luna Decisions is fourth of six on accuracy, behind Surogate Rune 26B-A4B at 79.7%.
The latency row mixes two harnesses. Microsoft timed its own model through Foundry, in the same region, with a 95th percentile of 125 ms. The other models’ figures are JevBench medians. Benchmark Heaven’s data gives OpenAI’s API a median of 298 ms and a 95th percentile of 399 ms over 598 requests. Microsoft-Decision-1 is not on JevBench as of 10 October 2026, so no independent latency figure exists for it yet.
OpenAI publishes no benchmark of its own, no latency in milliseconds and no calibration data. Its guide says the API returns answers about 10 times faster than the Responses API.
Price, context and hosting
Microsoft-Decision-1 is cheaper per token: $0.042 per million input tokens against $0.10. Neither charges for output. OpenAI adds that regional processing premiums and long-context multipliers still apply. The Foundry model card says the provider has not supplied pricing, so Microsoft’s launch post is the source for its price. OpenRouter lists both at their makers’ prices, with Microsoft’s model served by Azure.
The context gap is large. Microsoft-Decision-1 takes 32,768 tokens of text. OpenRouter lists OpenAI’s API at 1,050,000 tokens, the same window as GPT-6 Luna. Only OpenAI’s reads images.
OpenAI supports Zero Data Retention and HIPAA use for eligible customers, with data residency in the United States and Europe. Microsoft sells its model in Foundry as a Direct from Azure model. The Foundry model card says the model is tuned for English and should not be the only basis for decisions about people, such as hiring or credit.
Microsoft plans to rebase on OpenAI models
The launch post says Microsoft will “soon rebase it on other models, including Microsoft AI (MAI) and OpenAI.” Today Microsoft-Decision-1 is a Qwen3.5-9B model. Microsoft gives no date and no OpenAI model name. If it ships, a later Microsoft-Decision-1 could sit on an OpenAI base model while competing with OpenAI’s own Decisions API. Until then, the two run on unrelated base models from different companies.
How both compare with Jev
Jev is the reference model on this site. It costs $0.042 per million input tokens, the same as Microsoft-Decision-1 and under half of OpenAI’s price. It reads text only, up to 64,000 tokens per request, and uses its own /v1/systemone format with TypeSafe’s Python and TypeScript SDKs. OpenAI’s format differs from it, and Microsoft has not said whether its model accepts it.
On speed, JevBench gives Jev 1.13.0 a median of 239 ms, a little under OpenAI’s 298 ms. Microsoft’s chart shows Jev at 240 ms against its own 85 ms, from different harnesses. Jev has no accuracy or calibration score in Microsoft’s charts. Every, which tested the OpenAI Decisions API in preview, found it beat Jev on a few early tests and trailed on its broader evals. Jev vs Microsoft-Decision-1 covers that pair in full.
When to use which
- Price matters most. Microsoft-Decision-1 costs $0.042 per million input tokens against OpenAI’s $0.10.
- You need images or long inputs. Only OpenAI’s API reads images, and its window is far larger than 32,768 tokens.
- You want to start coding today from published docs. OpenAI’s request format and SDKs are documented. For Microsoft’s model, OpenRouter’s Decisions endpoint is the documented route as of 10 October 2026.
- You need general availability. Microsoft-Decision-1 is GA. OpenAI’s API is a public beta, with GA expected in the coming weeks.
- You trust a vendor’s benchmark. Microsoft’s charts favour its own model on accuracy, calibration and speed. Run both on a sample of your own data before you pick.
See Jev alternatives for the other hosted and open options.
FAQ
Is GPT-6 Luna Decisions the same as the OpenAI Decisions API?
Yes. Microsoft’s launch post links the name GPT-6 Luna Decisions to OpenAI’s Decisions guide. OpenAI calls the product the Decisions API and runs it on gpt-6-luna, its only model. GPT-6 Luna Decisions is OpenRouter’s name for that API, which OpenRouter lists as openai/gpt-6-luna-decisions, served by OpenAI.
Is Microsoft-Decision-1 better than OpenAI’s Decisions API?
Microsoft’s own tests say so: 83.5% average accuracy against 79.4%, a calibration score of 92.2 against 89.9, and an 85 ms median against 300 ms. Microsoft ran those tests and took the OpenAI latency from JevBench, a different harness. No one else has compared the two yet, so test both on your own data.
Which is cheaper?
Microsoft-Decision-1, at $0.042 per million input tokens against $0.10 for OpenAI’s Decisions API. Neither charges for output tokens. OpenAI adds regional processing premiums and long-context multipliers. OpenRouter lists both at their makers’ prices as of 10 October 2026, with Microsoft’s model served by Azure.
Can I call Microsoft-Decision-1 with OpenAI’s request format?
Nothing documented says so. Microsoft has not published Microsoft-Decision-1’s request format or endpoint path as of 10 October 2026. OpenAI’s format uses input and a questions array at /v1/decisions. OpenRouter’s Decisions endpoint lists both models under one request shape, which is the documented route to Microsoft’s model for now.
Will Microsoft-Decision-1 run on an OpenAI model?
Microsoft says it will soon rebase the model on other models, including Microsoft AI’s MAI models and OpenAI’s. It gives no date and names no OpenAI model. Today Microsoft-Decision-1 is a post-trained Qwen3.5-9B from Alibaba. Check the model page for changes.
Examples
Microsoft-Decision-1: Our model for fast decision-making
Microsoft's launch post for Microsoft-Decision-1, a Qwen3.5-9B decision model on Foundry at $0.042 per million input tokens. It reports Microsoft's own accuracy, calibration and latency charts against Quyet-1.0-Large, GPT-6 Luna Decisions and Jev, plus internal use at Xbox Research and Copilot.
Satya Nadella (@satyanadella)454k viewsOpenMicrosoft-Decision-1: Our model for fast decision-making on commandline.microsoft.com