Baseten’s joey-1 makes AI decisions for a penny per million tokens

Baseten released joey-1, a 2B open-weights “ decision model ” that returns an answer and a probability rather than text. It costs $0.01 per million tokens and answers in about 20ms, Baseten says. It joins the larger mammoth-1 and elephant-1 launched last week.

Baseten has released joey-1, a small open AI model that answers yes-or-no and multiple-choice questions instead of writing text, the company said on Thursday . It costs $0.01 per million tokens.

Joey-1 has 2B parameters and joins the larger mammoth-1 and elephant-1, which Baseten’s research lab, Base Labs, released last week. Rather than writing a reply, the models return an answer with a probability score.

That makes them cheaper and faster for jobs such as routing support tickets, spotting spam or blocking harmful prompts. TypeSafe AI’s Jev , launched in September, works the same way.

Baseten says joey-1 answers a typical 2,000-token request in 20 milliseconds, about 2.5 times faster than elephant-1. Sorting 10M support tickets a month would cost about $50, against $500 on mammoth-1, it says.

The company calls joey-1 the strongest decision model with 3B or fewer parameters, but has not yet published the tests behind that claim. Its model card ranks joey-1 37th of 230 models on the Jev Decision Index, the top score for models under 4B parameters.

Joey-1 is built on Alibaba’s Qwen3.5-2B and was trained on 6.21 million decision examples. Microsoft built its own decision model on Qwen this month, priced at $0.042 per million tokens. Its weights are free on Hugging Face under the MIT licence.

Baseten, which helps companies run AI models, raised $1.5B at a valuation of up to $13B in June. In September it bought sandbox startup Blaxel .