The Essentials
Two months after the launch of Jev by TypeSafe AI, several players are offering their own decision models. Cloudflare, Microsoft, OpenAI, AWS, Google, and PostHog are exploring this space, which aims to produce structured and calibrated responses rather than generic text.
Architectures, supported modalities, and pricing remain quite diverse.
- A different use for these models: these systems answer predefined questions — selecting an option, assigning a score, determining whether a proposition is true or false, or locating an object in an image. They can handle multiple questions in parallel without the classic token generation step.
- Cloudflare bets on open weights: its three models Clef, Clef-flash, and Clef-omni derive from Qwen and are published under the Apache 2.0 license. They support up to 64 questions per query, with text, images, and, for Clef-omni, audio. The Clef-flash price has been lowered to $0.038 per million tokens.
- Microsoft reuses Qwen: Microsoft Decision-1 is based on Qwen3.5-9B. This closed model, accessible via Azure and Foundry, handles only text and provides 32,000 tokens of context. Microsoft cites internal uses for analyzing Xbox feedback, assessing Copilot, and retrieving context during incident response.
- OpenAI adapts an existing model: the Decisions API relies on GPT-6 Luna, with support for text, images, and JSON. It accepts up to 400 questions per call. The announced price is $0.10 per million tokens, reduced to $0.01 for cached inputs.
- AWS favors open-weight models: Strands Decider combines an initial version built on Qwen3.5-2B with four models based on Gemma 4. The latter also accept images, but their native context windows remain narrower than some competitors.
- Google weighs in on the market: Gemini Decision 1, launched on November 11, handles text, image, audio, video, and PDFs. Its context window reaches 256,000 tokens and it can process up to 1,000 questions per query. Its price is $0.10 per million tokens.
- PostHog keeps a degree of reasoning: its Jeeves model can trigger a reasoning step when confidence is insufficient. It is used, notably, for detecting personal data in session summaries, selecting charts, or choosing a model best suited to a query.
- The market remains difficult to compare: pricing, context size, modalities, licenses, and reasoning capabilities vary greatly. The choice will depend on the targeted tasks, integration constraints, and the true cost in usage beyond the token-per-million-price metric.
Two months ago: the startup TypeSafe AI, founded by a former OpenAI researcher, launched Jev.
This model sits at the intersection of LLMs and classifiers: it does not produce tokens but “calibrated decisions.” More precisely, it provides answers to four types of well-defined questions:
- Choose an option from a predefined list
- Assign a score on an ordinal scale
- Judge whether an item is true or false
- Draw bounding boxes
Each question has its primitive (respectively Choice, Score, Noul, and Box). They can be combined in a single call. Jev processes them in parallel, without task-specific training, with a 64k context window. Cost: $0.042 per million input tokens.
TypeSafe AI describes the model as a “System 1,” referencing the terminology used by Daniel Kahneman in his book Thinking, Fast and Slow. It denotes quick, intuitive judgment, as opposed to “System 2,” which involves slow, deliberate reasoning.
“Jev” refers to the British economist William Stanley Jevons (1835-1882), who theorized that drastically lowering the cost of a resource multiplies its usage — a reference used, for example, to illustrate the rebound effect of 5G.
Three Open-Weight Models at Cloudflare
Originally, Jev only managed the primitives Choice, Score, and Noul. The bulk of competitors that followed have adopted this trio.
Cloudflare was among the first to announce. In early October, it rolled out two models: Clef and Clef-flash. A few days later, Clef-omni joined the lineup. All derive from Qwen models (dense for Clef and Clef-flash, MoE for Clef-omni). Their weights are published under the Apache 2.0 license.
Clef and Clef-flash handle text, structured JSON, images, and video. Clef-omni adds audio. All accept up to 64 questions per query. On Workers AI, Clef and Clef-omni each feature 64k of context (expandable to 256k in self-hosted scenarios). That was also the initial case for Clef-flash. But Cloudflare reduced the window to 24k after observing that 99.76% of requests did not exceed this threshold. This allowed a more attractive pricing: $0.038 per million tokens, down from $0.09 previously.
Microsoft also relies on a Qwen LLM
Launched the same day as Clef-omni, Microsoft Decision-1 derives from the same LLM as Clef-flash (Qwen3.5-9B). It is a closed model, with a 32k context, available on Azure / Foundry (inference localized in the EU) and it handles text only. It costs the same as Jev: $0.042 per million tokens.
Microsoft provides a few internal use cases. Among them, the Xbox team to thematize user feedback and the Copilot team to measure answer quality. Another example: the built-in replanning function in the Microsoft Discovery “R&D agenting” platform. The model also supports context retrieval for incident response (logs, tickets, calls, etc.).
OpenAI Reframes GPT-6 Luna as a Decision Model
To compete with Jev, OpenAI did not build a brand-new model. It utilizes GPT-6 Luna via the Decisions API. Released to general availability on October 22, it accepts text, images, and JSON. Up to 400 questions per call. Price: $0.10 per million tokens, or $0.01 for cached inputs.
Beyond caching, the Decisions API enables inference localization within the EU and integrates with the real-time voice mode GPT-Live. OpenAI highlights several internal use cases, including transaction triage in the Personal Finance team and analyzing themes and tone of press coverage for the Communications team.
AWS Widely Bets on a Gemma Base
As with Cloudflare, AWS chose the open-weight path for its Strands Decider models. The first version is based on Qwen3.5-2B. “Small enough to experiment, big enough to accomplish meaningful tasks,” says Amazon’s cloud branch. The company also promotes fine-tuning, giving an indicative 11 hours of training on an RTX 3090.
Beyond this base, four more models were added, all built on Gemma 4, including one MoE (26B-A4B). Like Jev, these can accept images both as context and within prompts. AWS tends to keep the native context window under wraps, limited to 4k. The Gemma 4-based models have nevertheless been tested up to 32k.
Google Finally Launches Its Jev Too
Google waited until November 11 to unveil its response: Gemini Decision 1, which comes with its own API. It combines text, image, audio, and video modalities (including PDFs) with a context window of 256k, up to 1,000 questions per query, and the ability to adjust resolution. For now, there is no caching, no batch processing, and no Flex inference.
Access is currently available via Google AI Studio and the google-genai SDKs, priced at $0.10 per million tokens. Integrations are in progress with LangChain, OpenCode, and Vercel, among others. Google cites a few internal use cases across its products: Gemini Notebook to route each request to the right model and reasoning level, CC (Google Labs product) for emoji selection, and Google Search for the safety of MCP tooling.
PostHog Maintains a Dose of Reasoning
PostHog also bases Jeeves on Qwen as its underlying model. A notable feature is its ability to reason when confidence falls short (default threshold: score < 0.9). The code is MIT-licensed while the weights carry an Apache 2.0 license. The model weighs in at roughly twenty gigabytes in BF16 and about ten in FP8.
PostHog also highlights several internal applications. In particular, in its session replay tool: to identify the presence of personal data in summaries, and to determine whether a summary matches what the user is seeking. PostHog uses Jeeves with its Slack bot to optimize selecting the model that answers a given query. It is also used to inspect SQL queries and the initial results rows, helping choose the ideal chart.
