Baseten launches joey-1 decision model at one cent per million input tokens

Not every AI task needs a paragraph. A customer service system may need to decide whether a message belongs with billing or technical support. A moderation pipeline may need a yes-or-no answer, and a business workflow may need a probability that a request satisfies a policy. Baseten's new joey-1 model is designed to do those narrow jobs without first composing free-form text. The company launched the two-billion-parameter model on 29 October and offers access both through an API and downloadable model weights.[1]

Joey-1 accepts a state, which can include text or images, and one or more typed questions. The answer can be a choice among categories, a yes-or-no probability or an ordered score. Rather than composing a free-form response that another component has to parse, the model emits the structured decision and associated probabilities directly. An application can then set confidence thresholds or escalate ambiguous cases to a more expensive model or a person.[1]

Baseten reports a median end-to-end response time of 20 milliseconds for an average 2,000-token request at production scale. It compares that with 49 milliseconds for the larger nine-billion-parameter elephant-1 in its own serving tests. It also prices joey-1 at $0.01 per million input tokens through its model APIs, with no separately charged output token in the listed pricing. The model card identifies Qwen3.5-2B as the base and lists the model weights under the MIT licence.[1]

The arithmetic behind the one-cent headline

At the published price, five billion input tokens cost about $50 before any additional service or processing expenses. That is the token volume in the company's illustration of ten million support tickets containing roughly 500 input tokens each. Baseten compares $50 with $500 for the larger mammoth-1 at its listed rate. This is a pricing illustration, not proof that joey-1 classifies ten million real tickets with the accuracy a business needs. Actual input length, repeated context, provider conditions and escalation rates can change the operating bill.[1]

The model also supports multiple questions against a shared input state in one request. That changes the economics when a workflow needs to check several criteria on the same document. A longer state does not have to be resent and reprocessed in independent calls if the questions can be bundled under the supported API. The model card shows the System One endpoint and typed question structures; there is no need to invent a bespoke instruction format for each classifier.[1]

Baseten says joey-1 is the strongest decision model with three billion or fewer parameters in its internal comparisons, while the detailed tests behind some performance claims had not been released with the launch announcement. The public Jev Decision Index is one piece of comparative evidence, but results on a decision benchmark should not be mistaken for universally calibrated probabilities in every industry's data. An insurance approval classifier and a low-risk support-ticket router have different error costs, class distributions and escalation requirements.[1]

For buyers, the meaningful question is how many decisions can be accepted automatically at a required error rate—not simply the cost of one million tokens. A model that is fast and cheap but frequently uncertain may send many cases to a larger model; a model that is confidently wrong may impose far higher downstream costs. Calibration must be checked against the target population, and higher-stakes workflows may need explicit audit records and human override.[1]

Deployment choice also changes the available context. Baseten lists a 16,000-token window on Model APIs and up to 262,000 tokens for dedicated deployments, which also support customised weights. Those are separate configurations: the inexpensive shared API does not automatically include the largest context window.[1]

Report an error or suggest a correction