Latest in AI
Baseten Joey-1: A 2B Decision Model at $0.01 per Million Input Tokens
Baseten's smallest decision model returns probabilities instead of prose, making routing and classification cheaper to put in front of larger models.
Baseten released joey-1 on October 29, a 2B decision model for classification, routing and scoring. It costs $0.01 per million input tokens through Baseten's API and is available as open weights. Instead of writing an answer, it returns probabilities for choices an application defines.
What joey-1 does differently
The release announcement describes joey-1 as a model post-trained from Qwen3.5-2B. Give it a state and typed questions, and it returns decisions that software can consume directly. Several questions about the same state share its input prefix, reducing repeated processing compared with separate requests.
This is a different job from asking a chatbot to explain a support ticket. An application might need to know whether the ticket concerns billing, whether it requests a refund and which urgency category applies. Those are bounded decisions. Generating several paragraphs and parsing the answer adds work that the application may not need.
Joey-1 supports yes/no questions, choices and ordered scores. Its larger siblings, elephant-1 and mammoth-1, remain options when the decision requires more capability. The new release adds a smaller first step to that family rather than replacing it.
Joey-1 pricing and API access
The Baseten model listing specifies the following:
| Item | Published detail |
|---|---|
| Model ID | decision-models/baseten-joey-2b-latest |
| Endpoint | /v1/systemone |
| Input price | $0.01 per million tokens |
| Cached input price | $0.00 per million tokens |
| Output price | $0.00 per million tokens |
| Licence | MIT |
The zero output rate reflects a decision-model product, not free text generation. Applications receive structured decisions rather than a normal chat response. An integration should use the documented request and response format rather than assume a chat-completions client can be swapped over unchanged.
At the listed input rate, processing one billion uncached input tokens would cost $10 in model input charges. That calculation excludes the rest of an application's infrastructure and says nothing about how accurate its decisions will be. Cheap inference is useful only if the resulting errors remain acceptable.
Context window, weights and benchmark caveats
The official model card specifies a maximum sequence length of 16,384 tokens, text and image inputs, and a self-hosting setup using one CUDA GPU. Its serving instructions say requests exceeding the trained length limit are refused rather than silently truncated. Baseten separately advertises a 16k Model API window and a 262k window for dedicated deployments in its launch post. Check the deployment you are actually using rather than treating those limits as interchangeable.
Base Labs reports a score of 63.30 on the public split of the Jev Decision Index 0.4, ranking joey-1 first among models under 4B in that evaluation. Its reported elephant-1 score is 72.24. Those are useful size-versus-capability signals, not evidence that joey-1 is equally reliable across every business classification task.
The card also reports 62.9 on the team's replica of the Artificial Analysis Decision Index. A replica result should not be presented as an independently published Artificial Analysis score.
Where a small decision model makes sense
A practical design is to give the small model a narrow decision and define an escalation path for uncertain or consequential cases. For support routing, that might mean assigning routine tickets automatically while passing ambiguous ones to a larger model or a person.
The threshold should come from testing representative examples, including the cases that are easy to confuse. A probability is useful input to a policy, but it is not permission to take an action. Classifying a message as a refund request should not, by itself, authorise moving money.
What becomes more interesting at this price is the placement of decision models throughout an application. Teams can consider checking more intermediate steps without paying for a full conversational response each time. The engineering challenge moves toward defining good questions, measuring errors and deciding which mistakes are worth escalating.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical teams explain model releases, developer workflows and product decisions without turning the engineering into marketing language.