The Pulse
Baseten Releases Joey-1, a 2B Decision Model Priced at $0.01 per Million Input Tokens
Baseten announced joey-1, a 2-billion-parameter model for decision tasks, priced at $0.01 per million input tokens through its Model APIs. Its model card reports a 63.30 score on the public split of the Jev Decision Index 0.4 and offers the
AI.info Team · Written with AI · fact-checked ·
Baseten announced joey-1 on October 29, 2026, a 2-billion-parameter model designed to make decisions without generating a written response. The company says the model returns a decision and its confidence in one forward pass, targeting workloads such as sorting support tickets or checking requests against approval rules.
Joey-1 returns probabilities, not prose
Applications provide joey-1 with a state—text, optionally accompanied by images—and a set of questions. The model card describes yes-or-no questions, choices among named options and ordered scoring, returning probabilities for available answers. That gives software a structured result directly, rather than asking it to generate text that another system must interpret.
Baseten presents the format as useful when decisions happen frequently or sit in the path of a user-facing response, where the cost and delay of each call matter. The company’s announcement defines the model’s output this way:
“A decision model returns a label (a decision) and a confidence score for that label in one forward pass.”
Max Kirkby and Charles O’Neill, authors of Baseten’s announcement
Baseten puts speed and price up front
Baseten reports a P50 end-to-end latency of 20 milliseconds for an average 2,000-token request. It compares that with 49 milliseconds for its 9B elephant-1 model and says joey-1 costs $0.01 per million tokens. The Hugging Face model card specifies that the API price is per million input tokens. Baseten says that makes joey-1 one-tenth the price of 27B mammoth-1.
The company illustrates the price difference with a hypothetical workload of 10 million support tickets per month, each using about 500 tokens. Baseten estimates that volume at roughly $50 monthly on joey-1, compared with $500 on mammoth-1. Those are company calculations; actual usage depends on the requests sent.
Baseten describes joey-1 as a first-line model for keeping workloads fast and cost-efficient, with elephant-1 or mammoth-1 available as escalation options when joey-1 is uncertain. That makes the confidence score operationally relevant: it can help an application decide whether to use the smaller model’s result or send the request to a larger one.
The benchmark shows a narrower strength than the launch claim
Baseten says its internal testing found joey-1 to be its strongest decision model with 3B or fewer parameters, while noting that details of those tests were still to come. On the public split of the Jev Decision Index 0.4, the Hugging Face model card reports a score of 63.30, ranking joey-1 37th among 230 models. In the same table, 9B elephant-1 scores 72.24; joey-1 also scores higher than the listed Cloudflare Clef, Jev 1.14 and Perplexity Decider Flash v1.2 results.
The card says joey-1 trails elephant-1 in every area on that split and beats Jev 1.14 in four of five. The results distinguish the smaller model’s lower-cost, latency-sensitive use case from a claim of leading the full family on benchmark performance.
Available through Baseten and as open weights
Baseten offers joey-1 through Model APIs, with a 16,000-token context window, and publishes the model weights on Hugging Face. The model is post-trained from Qwen3.5-2B by Baseten’s Base Labs, using the same data and recipe as elephant-1 and mammoth-1. Baseten also offers dedicated deployments with a 262,000-token context window and says joey-1 can be fine-tuned through its Training Jobs.
For developers, joey-1 is intended for a defined choice, yes-or-no judgment or score, rather than an open-ended answer. Its hosted API is priced at $0.01 per million input tokens, and its weights are available on Hugging Face.