Baseten's $0.10 Model Returns Probabilities Instead of Paragraphs

Baseten says bison-1 tops 19 decision models, at $0.10 per million input tokens and nothing for output.

On November 20, Baseten and Base Labs released bison-1, a decision model built on Google's Gemma 4 31B that scores text or image questions with option probabilities instead of writing free-form text. Jon's agents already sort and check with Jev, a decision model of the same kind, so bison-1 is a direct rival. Through OpenRouter, it runs at $0.10 per million input tokens and $0 output. It answers in probabilities rather than prose: a choice or a score with the odds behind it, or for a yes-or-no question, the chance of yes.

The hosts are AI characters.

Listen to this story
1:41

Baseten's numbers make the pitch sound simple. Its launch post says bison-1 scores 76 on Artificial Analysis Decision Index v2.1, placing first among 19 decision models ahead of d1.2-ultra at 75 and Baseten's own mammoth-1 at 71. Baseten also says bison-1 posts the index's lowest Brier score—a measure of how closely a model's stated confidence matches how often it is right, where lower is better—at 0.14 against d1.2-ultra's 0.15.

First of 19 is the maker's trophy from one index, not proof that bison-1 wins every judgment call. A Marktechpost report on X puts bison-1 in 11th place with a score of 60 on the knowledge-heavy AA-Omniscience-Decision part of that test, a warning for routers that lean on broad facts. Pratik Gandhi reported on X that bison-1 scored 221 of 231 on JevBench on his run, compared with 218 of 231 from Baseten's hosted API on the same questions. JevBench is a different test from the Decision Index.

Still, the release gives builders options beyond one hosted endpoint. Teams that want to self-host can download bison-1's open weights on Hugging Face, where Google's Gemma 4 31B IT base sits under the Apache License 2.0 and Baseten's changes use the MIT License. Baseten also launched a cheaper sibling, groundhog-1, at $0.035 per million tokens on Baseten Model APIs. Baseten says it expects groundhog-1 to match mammoth-1 on the official index, though Marktechpost reports it has no official index score yet.

Before wiring bison-1 into a live router, price the workload at $0.10 per million input tokens and check the context window—how much text the model takes in at once—on the exact endpoint that will serve traffic. Baseten's model library and Hugging Face card list 131,072 tokens, which its blog labels 128k, eight times its earlier models. OpenRouter's page, though, prints 33K right beside the price. Verify which ceiling the live endpoint actually enforces before sending a long policy through it, and test real, knowledge-heavy routing questions before retiring a general model.

Sources

Comments

No comments yet.

Sign in to comment. You'll come back to this story.