Microsoft joins AI's new decision-model rush as TypeSafe raises $870M
A new class of AI that skips writing text and just returns probabilities for preset options is exploding. TypeSafe AI kicked it off September 15 with Jev, and OpenAI, Cloudflare, AWS, Perplexity and Microsoft have all shipped rivals in under a month. On October 9, TypeSafe stacked another $870 million on top, at a reported $7.5 billion valuation.
- Original source: note.com.
- The source published it on Nov 15, 2026 at 4:50 AM.
- No second outlet has reported it yet. Treat it as a single-source claim.
Microsoft opened 'Microsoft-Decision-1' as a public preview on its Foundry AI development platform on October 9. The same day, TypeSafe AI — the U.S. startup that started the whole category — announced a fresh $870 million raise (about 138.2 billion yen), with its valuation reported at $7.5 billion (roughly 1.1916 trillion yen). That follows the $40 million round TypeSafe disclosed on September 15 alongside Jev, its first model, which launched in early access via a waiting list.
The copycats moved fast. Vercel started offering Jev through its AI connection service on September 16, and says about 13% of paid teams used it within the first 24 hours, making it the fastest-spreading model on the service, according to various news outlets. OpenAI announced its Decisions API on September 29, pushed it to public beta on October 6 and made it generally available October 22. October 1 brought a three-way pileup: Cloudflare's Clef and the smaller Clef-flash with weights released under Apache 2.0, AWS's Strands Decider 2B with weights, training data and training procedures included so it runs on a personal computer, and Perplexity's pplx-decider, per various news outlets. In just over three weeks, at least six companies had shipped models of the same type.
TypeSafe calls the category a 'System One model,' borrowing from the psychology split between fast, intuitive System 1 thinking and slow System 2 reasoning. Using one takes three pieces: a state (the inquiry or transaction record to judge), a question, and a list of answer candidates. A decision model reads the input once, picks from the options and returns a probability for each — up to 255 candidates in Jev's case. Question formats are near-identical across vendors: yes/no, multiple choice, and ordered tiers like light, medium or heavy. The model never writes out an explanation, which is why a judgment takes 70 to 500 milliseconds and costs $0.042 per 1 million input tokens at TypeSafe, with output free.
The humans still own the call. The program wrapping the model sets the thresholds — route to billing if the probability clears 90%, otherwise queue it for a person. Because there is no text to generate, many vendors skip output charges entirely and price only on input, with the source putting the spread at roughly 1.6 to 38 yen per 1 million input tokens, converted at 158.88 yen to the dollar. TypeSafe's Jev sits at the low end at $0.042; Perplexity's pplx-decider-flash comes in at $0.01, about 1.59 yen, with the 27B pplx-decider at $0.02.
Three of the models share DNA. Cloudflare's Clef, Microsoft's Decision-1 and AWS's Strands Decider 2B are all built on Qwen, the text-generating model from China's Alibaba, retrained to spit out scores instead of sentences. Developers argue the models 'do not generate nonsense' because they can't return answers outside the candidates or the format. External verification pushes back: what's guaranteed is the shape of the answer, not whether the judgment is right — it can still pick the wrong option. And the developers themselves say Japanese handling lags English, so users are expected to check the model against their own data.
TopicsTypeSafe AI · Jev · Microsoft · OpenAI · Cloudflare · AWS · Perplexity · Alibaba
Source: note.com
Written by AI from verified reporting. Not financial advice.

