TypeSafe launched Jev on September 15, 2026, and the decision model soon became one of the most visible new tools in AI. Jev is not a chatbot. It returns a label, a score, or a yes/no probability, giving developers a focused way to choose between defined outcomes.
That narrow design has created a new market in a matter of weeks. At least 20 Jev alternatives shipped in about four weeks, ranging from small open models to hosted systems from major AI companies. Jev itself became Vercel’s fastest-adopted AI model in its first two weeks, with 35% of paid teams using it.
“Jev has become the fastest adopted AI model ever — at least by some measures,” Russell Brandom said.
Why Jev Started the Race
TypeSafe has not disclosed Jev’s size, but the model offers a 64k-token context per request. TypeSafe’s documentation lists a price of $0.042 per 1M input tokens, with output free. Encoders such as Von and Laya run on CPU, giving the system a place in decision tasks that do not need a chatbot interface.
Jev is TypeSafe’s first “System One” model. In TypeSafe’s workflow evaluations, it scored 67.8% agreement and took 0.4 seconds per case. Its JevBench score is 72.07%, which gives the model a clear reference point for the alternatives that followed.
“Published in limited release on September 15, Jev is designed as a lightweight tool for choosing between a few predetermined outputs.”
The adoption numbers explain why so many competitors arrived fast. Vercel reported that Jev reached 35% of paid teams within two weeks, the fastest adoption rate ever for a model in its first two weeks by Vercel’s measure. Vercel also reported that open source models handle more than 70% of tokens processed through its gateway, while accounting for nearly 25% of total spend. Spending on frontier labs has grown tenfold.
Cloudflare, Liquid AI, and OpenAI Take Different Paths
Cloudflare announced Clef on October 1, 2026. Clef is a 27B model built on Qwen3.8-27B with a vision encoder, while Clef-flash is a 9B model built on Qwen3.5-9B. Cloudflare calls the API fully Jev-compatible, giving developers a direct alternative to Jev’s interface.
Clef leads Jev on BANKING77, scoring 94.20 F1 against Jev’s 79.74. Jev still leads on GPQA Diamond, where it scores 78.3 compared with Clef’s 48.0. The comparison shows that no single score settles the question of which decision model is best.
Liquid AI’s d1 family includes hosted d1 and d1-ultra models, along with two open models. The d1-3B model has 3.1B parameters and runs on LFM2.5-VL-3B. The d1-omni-600M model has 587M parameters and adds audio clips up to 30 seconds.
Liquid reports a 48.57 score for d1-3B on Decision Index 0.2.1, along with 8 ms latency on an RTX 4090. Those figures place d1 in the part of the market focused on small models and short response times.
OpenAI released the Decisions API in public beta on October 6, 2026. The API runs only on gpt-6-luna, accepts text plus base64 images, and supports three question types: predicate, choice, and score. Pricing is $0.10 per 1M input tokens, with output not billed.
More Choices for Narrower Decision Tasks
Perplexity’s pplx-decider-v1-27b is a Qwen3.8-27B fine-tune released under Apache-2.0. It reads text, JSON, and up to eight images per question, while its API supports 128 questions and 255 options per request. The listed price is $0.02 per 1M input tokens.
meraGPT announced Decider 1 on September 22, 2026. It is a hosted, closed-weight, text-only model that supports up to 10 Choice labels and a 4,096-token context. Decider 1 lists a price of $0.04 per 1M input tokens.
Respan’s Span-01 targets a different job. It is a behavior-scoring model rather than a general decision model, returning present, absent, or not_observable probabilities. Span-01 costs $0.02 per 1M input tokens, while Span-01 Lite is free. Respan claims 0.843 overall F1 on its Behavior Benchmark, with a lower F1 score of 0.806 also listed.
Standard Thinking Standard One ships as open and hosted 3B and 8B models. Standard One 8B scores 54.95% on JevBench, below Jev’s 72.07%, but leads on realistic transfer with 90.83% against Jev’s 86.67%.
The full field stretches from GLiNER2.5-multi-Decide at 287M parameters to Mapika Decider at 35B-A3B. GPT-6 Sol scores 0.885 on its benchmark, adding another figure to a market that now measures decisions across different tests.
Jev’s early lead has not stopped the competition. Instead, its launch has produced a menu of models with different prices, input types, context limits, speeds, and benchmark strengths. For developers choosing a decision system, the key question is no longer whether an alternative exists. It is which type of decision the model needs to make.
Based on
