Microsoft Decision-1 Undercuts OpenAI Decision Pricing by 58%

A cheap model for a narrow job

Microsoft has joined the crowded new market for "decision models." On October 9 it released Microsoft-Decision-1, a model that does not write prose. Instead, it gives a probability to each answer in a fixed list of options.11 The model takes a piece of text and a closed set of choices, such as yes or no, multiple choice, or a rating scale. In a single pass it returns calibrated odds for each one, and it can also grade other AI systems' answers against a rubric.1112 It costs $0.042 per million input tokens, output is free, and it is available through the Microsoft Foundry catalog to anyone with an Azure account.11

Some headlines say Microsoft "obliterated" its rivals at half the price. That claim comes from one comparison. OpenAI's Decisions API charges $0.10 per million input tokens, so on a per-token basis Microsoft's model is about 58% cheaper.12 The rest of the pricing picture is less dramatic. Cloudflare's Clef-flash charges $0.038 per million input tokens, which is slightly less than Microsoft.12 TypeSafe, whose Jev model started this category, charges exactly the same $0.042, which one outlet described as Microsoft copying the price.11 Decision-1 is a strong, cheap entry in a market that is already cheap. It undercuts OpenAI but is not the lowest price available.

What Microsoft says the model can do

The performance figures come from Microsoft and should be read that way. The company says Decision-1 had the highest accuracy in an evaluation of 36 benchmarks with nearly 150,000 held-out questions. The tests covered routing, ranking, multilingual tasks, reasoning and safety.12 One report puts the accuracy figure at 83.5%. On calibration, meaning how closely the model's stated probabilities match real outcomes, it scored 92.2%. That was second to Quyet-1.0-Large, so Microsoft did not lead on every measure.11

Speed is the main selling point. Microsoft says Decision-1's median latency is about 35 times faster than GPT-6 Sol. Jev and H2O-Lightning-4B are 2.8 and 2.5 times slower, respectively.12 Another report adds that Microsoft measured it as 4.5 times faster than Quyet-1.0-Large.19 Even Microsoft notes that the GPT-6 Sol comparison is not like-for-like. Sol is a general-purpose model, so comparing it to a specialist doing classification will always favor the specialist.12

The internal deployment numbers are where the claims become most concrete. According to the company, Xbox Research classified more than 10,000 pieces of user feedback with quality competitive with GPT-6 Sol, at more than 14 times the speed and one two-hundredth of the cost.1119 The Copilot team reported quality close to a model in OpenAI's Luna family at roughly 100 times the speed. Microsoft Discovery reported scoring consistency 46 times higher than LLM-based grading.11 Microsoft's own example explains why latency matters: if an agent makes 20 sequential decisions at 100 milliseconds each, the workflow gets two seconds slower.11

The Qwen detail

The biggest strategic surprise is the model's base. Decision-1 is a post-trained version of Qwen3.5-9B, Alibaba's 9-billion-parameter open-weight model, not a Microsoft model.1112 Microsoft has said it will soon rebuild Decision-1 on other bases, including its own MAI models and models from OpenAI, but it has not said why it started with Qwen.11

The practical reasoning is easy to follow. For a classification task, a small open model is cheap to fine-tune and can be served without large GPU clusters.11 It still looks odd for a company that has its own in-house model line and privileged access to OpenAI's models. My reading is that this matters more than the price headline. In this category, the price and the engineering decide the choice, not brand loyalty, and Microsoft picked whichever base got it to market fastest. There is a downside for customers. A new base model means new behavior, so teams that integrate Decision-1 now will need to re-test accuracy when Microsoft swaps the base.11

A category that appeared in weeks

Decision-1 is a late entry. TypeSafe's Jev made single-pass decision scoring popular in mid-September. Since then, by one count, more than a hundred similar models have appeared, including OpenAI's Decisions API, Cloudflare's Clef and AWS's Strands Decider 2B.11 Perplexity has also released a similar product.12 Microsoft's launch post compares Decision-1 with six of these rivals, including Strands Decider 2B, GPT-6 Luna Decisions and Surogate Rune 26B-A4B.11

The architectures differ. Decision-1 is a transformer. Jev has been described as a non-transformer design that its maker markets as a "System One" model, a term Microsoft does not use.11 Published latency figures are also hard to compare. TypeSafe quotes 70 to 500 milliseconds for Jev, while Cloudflare measured a median of 524 milliseconds, and those tests used different hardware, text lengths and methods.11 That is a reason to be skeptical of the "2.8 times faster" claim. Microsoft ran the test on its own setup.

Part of a broader price war

Decision-1 lands during a run of price cuts across the industry. In late September, OpenAI released GPT-6 Sol and GPT-6 Luna with API prices half those of their predecessors. Anthropic launched Claude Opus 5.5, which it says costs about 40% less to run than Opus 5 on typical workloads.1517 Analysts linked those cuts to corporate customers trying to reduce token spending and moving toward cheaper open-source models.17 On October 7, Anthropic released Claude Haiku 5.5, which it says costs about 75% less on average than Haiku 4.5.1320 Open-weight options also keep growing, including Reflection AI's Beam, which is pitched as a cheaper US alternative to the leading Chinese open models.20

That context makes "half the price" less striking. Halving prices has become the normal way to launch a model this season. Decision-1 stands out because it pairs a low price with a narrow task. It is built for the high-volume routing and grading work that agents do between their more expensive reasoning steps.

Microsoft's own model lineup is uneven

The same week shows the limits of Microsoft's in-house work. On October 1, its MAI group released three audio models. MAI-Transcribe-2-Streaming, the company's first real-time transcription model, posted a 2.5% final word-error rate with about 0.13 seconds to final output on Artificial Analysis's streaming leaderboard, ranking first among 38 models.78 Microsoft AI chief Mustafa Suleyman said the voice models were 55% faster and 60% cheaper than ElevenLabs.4 Those figures apply to MAI-Voice-2.1-Flash, a text-to-speech model priced at $15 per million characters.15

The transcription model, however, was not priced to undercut anyone. Its introductory price of $9 per 1,000 minutes matches Gemini 3.5 Transcribe Live and sits above ElevenLabs and Deepgram, both at $6.50.8 One analysis called the launch a play on accuracy and latency at matching prices, not a price war.9 Another described the pricing as premium.7 Some coverage also pointed out that the leaderboard ranking was reported by Microsoft and had not yet been confirmed independently.6

At the frontier, Microsoft is further behind. Benchmark trackers show its reasoning model, MAI-Thinking-1, scoring 52.8% on SWE-bench Pro, far below the leading score of 89.9%.10 CNBC reported that Suleyman's group has released coding, image, reasoning and speech models but has not matched Anthropic or OpenAI on independent leaderboards for raw capability.14

What it adds up to

Put together, the launches show a clear pattern. Microsoft is not winning the race for the most capable general-purpose model. It is winning in narrow, high-volume tasks such as transcription, voice and decision scoring, where speed and price per call matter more than general intelligence. This fits the strategy CNBC describes: Satya Nadella is moving Microsoft toward consumption-based pricing and selling tokens in bulk.14 Cheap specialist models that sit inside Copilot and Azure workflows match that goal well.

Whether Decision-1 "obliterated" anyone is still unproven. Its benchmarks are self-reported, its price is matched or beaten by smaller rivals, and its underlying model will change. Teams thinking about switching should test it on their own data before relying on Microsoft's 36-benchmark claim.11