TypeSafe raises $870M for Jev, a model that returns decisions

TypeSafe AI, the company behind the decision model Jev, has raised $870 million at a $7.5 billion valuation [1][2]. Andreessen Horowitz led the round, with Sequoia Capital and existing investor DCVC taking part [1]. The size of the round matters less to most readers than what Jev is: a model that does not write text at all, but answers narrow, typed questions for software to act on, at $0.042 per million input tokens with free output [3]. On the same day, Microsoft released a model built for the same job at the same price [4][5], and OpenAI’s own Decisions API is already in public beta [6]. If you pay a chat model to sort, route or grade things, a cheaper class of tool now exists for that work.

What was announced

TypeSafe announced the round in a short blog post it titled “TypeSafe A raises Series AI” [1]. It calls the round “a really big series A”, led by Andreessen Horowitz, with participation from Sequoia Capital, existing investor DCVC and a group of angel investors [1]. The amount and valuation come at the end of the post: $870 million at a $7.5 billion valuation, with Martin Casado joining the board [1]. TechCrunch reports the same figures and investors [2].

The company says little about how it will spend the money. Its post promises developers “even more machine-native models” and the infrastructure to build software around them, and promises businesses the enterprise features they have asked for [1]. It says a third of the Fortune 500 are using Jev and that it has saved customers millions of dollars in production [1]. Neither claim comes with names or figures that an outsider can check. The post gives no revenue number.

TechCrunch adds the background. Jev was released on September 15 and spread quickly [2]. TypeSafe was co-founded in 2024 by Diogo Almeida, previously a researcher at OpenAI, with former Meta research engineer Sasha Sheng and engineer Erik Gafni [2]. According to [2], Jev is based on a transformer architecture but is not a large language model: it does not output text, and instead produces probabilities, which the company calls “calibrated decisions”. TypeSafe claims it works significantly faster and uses far fewer tokens than LLMs, and positions it for automating tasks rather than generating text or code [2].

TypeSafe’s documentation explains what that means in practice [7]. You send Jev a piece of input, which TypeSafe calls the state, and one or more typed questions. There are three kinds. A Choice picks one option from a list and returns the choice, a probability for each option and a confidence value. A Score rates the state against a rubric and returns a score, probabilities and confidence. A Noul asks whether a statement is true and returns a probability between 0 and 1 [7]. All three kinds can be mixed in one API call, and each question is evaluated in parallel and in isolation against the same state [7]. TypeSafe tells developers to keep each question narrow, and to break a broad judgment such as “rate this startup pitch” into separate questions whose answers are combined in code [7].

The model page gives the terms [3]. The current model, Jev 1.13, costs $42 per billion input tokens, which is $0.042 per million, and output tokens are free. The context limit is 64k tokens per request, with 32k tokens for the state plus the longest question. Rate limits are 100K tokens per second and 80 requests per second, and TypeSafe warns that these are adjusting dynamically and can change without notice while it serves heavy demand [3]. Jev accepts text only; images, audio and video must be converted to text first. English is its primary language, and TypeSafe says other languages are handled but not equally well [3]. The page also says Jev is not trained on customer requests or responses, and that zero data retention is available to enterprise customers [3].

The same day brought a direct rival. Microsoft released Microsoft-Decision-1 in Microsoft Foundry and through OpenRouter, for routing, classification, prioritisation, verification and workflow control [4]. Microsoft says it post-trained Qwen3.5-9B to build it [4]. Its input price is also $0.042 per million tokens, and output is also free [4][5]. OpenRouter lists a 32,768-token context window [5]. Microsoft’s post carries an editor’s note saying it was updated to add benchmarks for Jev [4], so Microsoft is already benchmarking itself against Jev.

OpenAI is in the same market. Its Decisions API, in public beta with general availability expected in the coming weeks, returns a probability that a condition is true, a choice from a fixed set, or a score against a rubric [6]. It runs on gpt-6-luna, costs $0.10 per million input tokens, and has no output, cache-read or cache-write charges [6]. OpenAI says it is about 10x faster than its Responses API [6].

What changed, and for whom

The change is not the funding itself. It is that “decision model” is now a product category with at least three priced options from three companies, and two of them charge the same list price. For anyone who uses a chat model to make many small yes-or-no or pick-one calls, that is a reason to look again at where the tokens go.

The first group is people who build on the ChatGPT or Claude APIs. Many apps already use an LLM for jobs that are really decisions: is this email a complaint, which team should get this ticket, does this reply promise a refund. A chat model answers these in text, which then has to be parsed. A decision model answers with a typed value and a probability your code can act on directly [6][7]. Microsoft’s documentation gives examples such as triaging support tickets and evaluating model output, and says the model returns no written rationale [8]. That is a cost and reliability argument, not a quality guarantee: you still have to check the answers.

The second group is people running agents. Microsoft pitches its model for agent controls, such as deciding whether an agent’s next step should continue, stop, retry or be handed to a human, and for grading AI responses and agent actions against a rubric [4]. Those checks happen many times per task, so their price and speed add up. Microsoft claims Microsoft-Decision-1 is 35 times quicker than GPT-6 Sol in its own benchmarking [4]. That figure is the vendor’s, not an independent test.

The third group is people who buy through Microsoft or OpenAI already. Microsoft-Decision-1 sits in Foundry, and the Decisions API sits in OpenAI’s platform [4][6]. Jev is served through TypeSafe’s own API [3].

Who it is for — and not

Decision models are for steps where the answer is a fixed set of options. Routing, tagging, filtering, scoring and yes-or-no checks fit. TypeSafe’s documentation describes each question as a gut-check determination that a knowledgeable person could make in a few seconds given the right context [7]. If your step looks like that, a decision model is worth testing.

They are not for writing. Jev does not output text [2], and Microsoft-Decision-1 does not generate a free-form response or written rationale [8]. If a person needs to read the output, or you need an explanation for an audit trail, a chat model still has to do that part.

They are not for long or mixed inputs without preparation. Jev takes text only, within a 64k-token budget per request [3], and OpenRouter lists Microsoft-Decision-1 at 32,768 tokens of context [5]. Long documents, images or audio need to be cut down or converted first, except on OpenAI’s API, which accepts text and images [6].

They are not proven on your data. Every speed and accuracy figure in this story comes from the vendors themselves. TypeSafe’s customer claims come without names [1]. TypeSafe also warns that Jev’s accuracy is best in English [3]. The probability each model returns is only useful if it is calibrated for your kind of input, which you can only learn by testing against examples you have already labelled.

And they are not yet stable products. TypeSafe says its rate limits can change without notice [3], and OpenAI’s API is still in beta [6]. Building a critical path on any of them today means planning for a fallback, the same way you would for a new chat model.

This page will be updated if TypeSafe, Microsoft or OpenAI change prices or rate limits, or when OpenAI’s Decisions API leaves beta.

∴ Yes-or-no and pick-one calls no longer need a chat model; test a decision model on your cheapest repetitive step.