Not every valuable AI model generates text. Since mid-September, a new category has been taking shape around models that don't write anything at all — they just take a fixed list of possible answers and return a calibrated probability for each one, built for jobs like routing a request, classifying a document, or deciding whether an AI agent's output passed a check. We covered the first of these, TypeSafe's Jev, when it launched a browser agent that finished tasks in about 7 seconds last month. On October 9, 2026, the category had its first real market test: TypeSafe raised a huge round at a huge valuation, and a hyperscaler showed up the same day pricing a rival model identically.
Quick summary: TypeSafe AI announced an $870 million Series A at a $7.5 billion valuation on October 9, 2026, led by Andreessen Horowitz with Sequoia Capital and DCVC participating — 24 days after its Jev model launched on September 15. The same day, Microsoft released Microsoft-Decision-1, a "decision-scoring" model built by post-training Alibaba's open-weight Qwen3.5-9B, priced at $0.042 per million input tokens with free output — an exact match of Jev's own pricing. Microsoft claims Decision-1 is up to 35x faster than its GPT-6 Sol and scores highest across 36 internal benchmarks (~150,000 questions) at 83.5% average accuracy — vendor-reported figures, not yet independently verified.
What a "decision model" actually is
A decision model doesn't generate open-ended text the way ChatGPT or Claude does. Give it a question and a fixed set of possible answers — "is this support ticket urgent, normal, or low-priority," "should this agent's output be approved or rejected," "which of these five tools should handle this request" — and it returns a calibrated probability for each option, not a written response. That narrower job means a much smaller model can do it well, and because there's no lengthy text generation, it can run dramatically faster and cheaper than routing the same decision through a full conversational LLM. For AI agent pipelines, where a single user request might trigger dozens of small classification and routing decisions before anything user-facing happens, that adds up fast.
TypeSafe's 24-day valuation jump
TypeSafe's Jev, a proprietary "System One" model, launched September 15, 2026, and quickly found a flagship use case: Browser Use's Jev Ultrafast agent, which cut a Google Flights search from roughly 9.45 seconds down to 7.1 seconds by having Jev decide the next browser action directly from a parsed page structure instead of reasoning over a screenshot. Twenty-four days later, on October 9, TypeSafe closed an $870 million Series A at a $7.5 billion valuation, led by Andreessen Horowitz with Sequoia Capital and DCVC also participating (Source: citybiz). Jev itself is priced at $0.042 per million input tokens with output free, runs on a 64,000-token context window, and is text-only. TypeSafe removed Jev's waitlist on September 21, 2026, and now grants new accounts $5 of free credit to try it (Source: eesel.ai).
Microsoft's answer: same price, different foundation
Also on October 9, Microsoft released Microsoft-Decision-1 — built not from scratch, but by post-training Alibaba's open-weight Qwen3.5-9B model into a decision-scoring system. It carries a 32,768-token context window (half of Jev's) and is priced at exactly $0.042 per million input tokens with output free — the identical rate TypeSafe charges for Jev (Source: TheNextWeb). Microsoft reports Decision-1 as up to 35 times faster than its own GPT-6 Sol and 4.5 times faster than a comparison model called Quyet-1.0-Large, with roughly 85-millisecond latency, and the highest accuracy across an internal suite of 36 benchmarks and about 150,000 questions, averaging 83.5%. Those are Microsoft's own reported figures, not yet reproduced by an outside evaluator, so they deserve the same "reportedly" treatment as any vendor benchmark claim this fresh. Decision-1 is available now through Microsoft Foundry and on OpenRouter; Microsoft says it plans to rebase future versions on other model families, including its own MAI models and OpenAI's.
| Model | Built On | Context Window | Price (in / out per 1M tok) | Backing |
|---|---|---|---|---|
| Jev (TypeSafe) | Proprietary "System One" | 64,000 tok | $0.042 / free | $870M Series A @ $7.5B (Oct 9, 2026) |
| Microsoft-Decision-1 | Qwen3.5-9B (post-trained) | 32,768 tok | $0.042 / free | Microsoft (in-house) |
Why a hyperscaler could undercut a $7.5B startup this fast
The uncomfortable question for TypeSafe's new investors is sitting right in that comparison table: Microsoft didn't need to train a foundation model from scratch to compete. It took an existing open-weight model that Alibaba already spent the money training, did a round of targeted post-training for the decision-scoring task, and shipped a product at an identical price point within weeks of Jev's funding news. If the core insight behind the "decision model" category turns out to be more about product packaging — how you frame the output, not a defensible architectural moat — then the bar to compete with a funded startup in this space may be lower than a $7.5 billion valuation implies.
If you're building an agent pipeline that needs cheap routing or classification, this is a good moment to actually benchmark both models against your own task rather than assume either vendor's published numbers transfer to your workload — pricing is now identical, so the real differentiator is which one is more accurate on the specific decisions you need made.
01What is a "decision model" in AI, exactly?
A model that scores a fixed list of possible answers with calibrated probabilities instead of generating open-ended text — used for routing, classification, verification, and workflow-control tasks inside larger AI agent pipelines, where it's much cheaper and faster than running the decision through a full conversational model.
02Is Microsoft-Decision-1 open-source like Qwen3.5-9B?
No. Microsoft built it by post-training Qwen3.5-9B, but Microsoft is not distributing the resulting weights — it's available only as an API through Microsoft Foundry and OpenRouter, not as a downloadable open-weight model.
03Why did TypeSafe reach a $7.5B valuation so fast?
Jev's launch in mid-September demonstrated a real, measurable speed advantage (roughly 9.45 seconds down to 7.1 seconds on a browser-agent task) for a genuinely new model category, which drew investor interest quickly. The $870 million Series A, led by Andreessen Horowitz with Sequoia Capital and DCVC, closed just 24 days after launch — an unusually fast valuation climb even by 2026 AI-funding standards.
Whether you're picking a decision model for an agent pipeline or just trying to keep up with how fast this corner of AI is moving, it helps to see more than one vendor's benchmark table. LumiChats puts 40-plus leading AI models in one place so you can compare how different providers actually perform on your own tasks, instead of taking a launch-week press release at face value.
