Microsoft Undercuts OpenAI 58% With Model on Alibaba's Qwen
Microsoft Undercuts OpenAI 58% With Model on Alibaba's Qwen
Microsoft has just picked a fight with its closest partner — and armed itself with a Chinese rival's technology. On Friday, the company launched Decision-1, a model that does exactly what OpenAI's brand-new GPT-6 Luna Decisions does, costs 58% less, and is built on Alibaba's Qwen3.5. If you were waiting for proof that the Microsoft–OpenAI partnership has quietly turned into a rivalry, here it is: priced in cents, benchmarked in public, and impossible to misunderstand.
Decision-1 is a decision model, not a chatbot. You hand it a piece of text and a fixed list of options — yes/no, multiple choice, a rating scale — and it returns a calibrated probability for each option in a single structured API call. No prose, no explanations, no chat. The output is JSON that software can act on immediately: above a confidence threshold the ticket routes itself, below it the case goes to a human. Microsoft is pitching it as the cheap, fast control layer for the age of AI agents.
There was no keynote. Nadella unveiled it in a post on X on a Friday afternoon, writing: "It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality." He added that Microsoft is "already testing it across Microsoft for everything from incident response and quality control to scientific discovery." The fuller announcement came as a Command Line blog post signed by Achint Srivastava, Microsoft's vice president of software engineering in the Office of the CTO.
The price is the story
Decision-1 costs $0.042 per million input tokens, and output tokens are free. OpenAI's Decisions API — the endpoint that runs GPT-6 Luna Decisions, launched in public beta on 6 October, barely three days before Microsoft's move — costs $0.10 per million input tokens. That is a 58% undercut on the exact same pricing unit, for the exact same job. Microsoft isn't competing with OpenAI's flagship models here. It's competing with the API OpenAI launched days ago, at a price OpenAI has no answer to yet.
The price carries one more sting: it matches the startup Jev's to the cent. TypeSafe's Jev charges $0.042 per million input tokens — the same figure. Jev is the TypeSafe AI product that lit up the decision-model category in mid-September after spreading through Vercel's AI Gateway. As one trade publication put it, Microsoft effectively copied the price. It did not invent this category; it saw one, waited three weeks, and showed up with the balance sheet.
Built on Alibaba's Qwen — yes, really
Here is the part that will raise eyebrows in Washington. Microsoft post-trained Qwen3.5-9B — an open-weight, nine-billion-parameter model from Alibaba, a Chinese lab — instead of using one of its own models or OpenAI's. The model runs a 32,768-token context window and accepts text only: no images, no audio, no video. Weights are closed, so you cannot download it and run it locally; it is a hosted API in Microsoft Foundry (generally available as a "Direct from Azure" model) and on OpenRouter.
We checked the public listing data: Decision-1 opens from the Microsoft Foundry catalog, requires an Azure account, and accepts text only. On OpenRouter it runs through a Decisions API endpoint — standard chat-completion SDKs will not work, a quirk developers will need to mind.
Microsoft also says it plans to rebase Decision-1 on other model bases soon, including its own MAI family and OpenAI's models — without explaining why it didn't start there. Practical answer: for a classification job, a cheap open 9-billion-parameter model is easy to train and cheap to serve. Strategic answer: Microsoft has in-house alternatives (its MAI line) and access to OpenAI's models, yet chose neither for version one. Cloudflare, incidentally, built its Clef-flash decision model on the very same Qwen3.5-9B base.
The benchmark brawl
Microsoft says Decision-1 ranked first in accuracy across 36 benchmark tests covering 147,137 questions, scoring 83.5% on average. The company's own table puts Quyet-1.0-Large at 81.9%, GPT-6 Luna Decisions at 79.4%, and H2O-Lightning-4B at 77.2%. Calibration — how closely the stated probabilities match reality — scores 92.2, second to Quyet's 93.1. On speed, Microsoft reports a median (p50) latency of 85 milliseconds, claiming 4.5 times the speed of Quyet-1.0-Large and 35 times GPT-6 Sol, plus decision-stability of 98.7% across eight kinds of text perturbation.
Now the caveats, because they matter. These figures come from Microsoft's own tests. None has been independently reproduced. Decision-1 does not appear on the JevBench leaderboard as of 10 October, and the speed figures compare Microsoft's Foundry-measured timings against competitor numbers derived from JevBench's adjusted medians — not the same harness. H2O.ai's model card disputes the comparison outright: it says JevBench's adjusted figure doubles measured time and adds 0.15 seconds, and that its model's measured median is 29 milliseconds, not the 210 milliseconds Microsoft charted. Cloudflare's open Clef models — built on the same Qwen base — were not included in Microsoft's comparison at all. The Register, which covered the launch, noted the same gap.
And then there is the internal-use pitch. Microsoft's Xbox Research team sorted more than 10,000 pieces of game feedback with Decision-1 and reported quality comparable to GPT-6 Sol at over 14 times the speed and roughly one-two-hundredth of the cost. The Copilot team found it competitive with GPT-5.6 Luna and 100 times faster. Microsoft Discovery claims scoring 46 times more consistent than an LLM-based approach. Big numbers — all self-reported, all from inside the building.
What it cannot do
The model card is blunt, and developers should read it before architecting anything. Decision-1 does not generate text, chat, translate, or summarise. It returns no explanations or rationales — just numbers. It must not be the sole basis for decisions about credit, employment, housing, healthcare, or legal rights. Applications must define their own options, thresholds, escalation paths, and human oversight. As Microsoft's own Achint Srivastava put it in a statement carried by several outlets, decision-scoring systems are "a new category aimed at overcoming latency bottlenecks in software automation" — the point is speed and structure, not intelligence.
The partnership is now a rivalry, priced to prove it
Let's be blunt. This is the most aggressive move Microsoft has made against its own partner since the two companies rewrote their deal in April 2025 and dropped the exclusivity clause. OpenAI launched a decisions endpoint on 6 October; Microsoft shipped a cheaper one, built on a competitor's open model, three days later, and published a table showing OpenAI finishing fourth. That is not how partners behave. It is how a former partner behaves when it has decided the margin in AI infrastructure belongs to whoever sells the cheapest dependable call.
Whether Microsoft can win this round is a separate question, and the honest answer is: on price and distribution, probably; on trust, not yet. The 83.5% is Microsoft grading its own homework. Until Decision-1 lands on an independent leaderboard — JevBench, the Decision Index, anywhere neutral — the accuracy claim is a marketing number. But the price is not a claim. It is a fact, it is live today, and it is the kind of fact that makes a CFO tell engineering to run the migration. OpenAI's moat was never the API; it is the models. Microsoft just proved the decision-API layer is a commodity, and commodities go to the cheapest credible supplier. Right now, uncomfortably for everyone else, that is Microsoft — running on Alibaba's weights.
What developers should actually do
If you route tickets, flag fraud, grade agent actions, or filter content with a general-purpose LLM today, take 200 real examples that are already labelled and run them through both your current setup and a decision model. Measure accuracy, latency, and cost per thousand calls on your own material. Microsoft's 36 benchmarks tell you where Decision-1 sits in Microsoft's ranking — not how it behaves on your data. That measurement takes an afternoon and is the only one that counts.
Two warnings before you commit. First, there is no local escape hatch: the weights are closed, the API shape is what Azure serves, and OpenRouter notes the weights update continually while the shape stays fixed. Second, Microsoft says the next version will be rebased on a different base model — which changes behaviour. Integrating today means re-measuring tomorrow. That migration risk does not show up on the price list.
And keep an eye on the independent boards. If Decision-1's benchmark lead survives contact with a neutral leaderboard, the decision-model wars are over before they began — Microsoft won on price alone. If it doesn't, this was a pricing announcement dressed as a benchmark victory. Either way, OpenAI has three days of head start and a 58% price gap to close. Awkward.
Decision models are the infrastructure nobody saw coming: Anthropic's Claude was caught steering wealthy flyers to tickets costing $198 more — exactly the kind of judgment call these small, fast scoring models are now meant to gate. Meanwhile, the agent era keeps producing stranger proofs of concept, like ChatGPT driving a Toyota to In-N-Out at 0.94 miles an hour — a reminder that the interesting question is no longer what AI can do, but which decisions we let it make alone, and at what price.
Update, Oct. 12: the two holes in Microsoft's victory lap
Two caveats have surfaced since Microsoft's Decision-1 launch post (Oct 9, by Achint Srivastava, VP of software engineering in the Office of the CTO). First, Microsoft's 36-benchmark comparison did not include Cloudflare's open Clef models — also built on Qwen — which means the “2.8x faster” and top-accuracy claims describe the models Microsoft chose to test, not the whole field (the-decoder, dataist.ai). Second, Microsoft itself says the Qwen base is temporary: the company expects to replace or supplement it with its own models or OpenAI's systems in a future version (TechDefused, MarkTechPost). That rebase promise cuts both ways — it answers the awkwardness of a US frontier firm shipping on a Chinese open-weight model, and it confirms the awkwardness is real. Meanwhile Jev, the TypeSafe's Jev startup that kicked off the decision-model trend in mid-September, is having “a rough few weeks” as the category it created gets swallowed by Big Tech (the-decoder).
>Sources: Microsoft Command Line launch post (Achint Srivastava); Satya Nadella on X; Microsoft Foundry catalog; The Register; MarkTechPost; TestingCatalog; The Decoder; OpenAI developer docs; CellCog; Vincent Vandeth's independent Decisions API testing.TechPulse USA — by the TechPulse USA newsroom. We cover the AI industry the way a newsroom should: every fact checked against multiple sources, every quote attributed, every image licensed and inspected. Got a tip or a correction? We read everything.
Comments
Post a Comment