Baseten releases Joey-1 for AI decisions at $0.01 per million input tokens

Baseten released joey-1 on October 29, 2026, giving developers a smaller AI model for decisions such as routing support tickets or flagging messages. The 2-billion-parameter model is available through Baseten Model APIs at $0.01 per million input tokens, with downloadable weights under the MIT license for those who want to run it themselves.

The release adds a speed-and-cost option to Base Labs’ existing decision-model family. Baseten positions joey-1 for high-volume work where response time matters most, rather than as a replacement for its larger models. It recommends those larger models for decisions requiring greater accuracy or broader general knowledge.

Choosing an answer, not composing one

A decision model takes an input and questions with predefined answers, then returns probabilities for the options. Joey-1 accepts text and images. Instead of composing a written response that an application must interpret, it produces a label and confidence score in one processing pass through the model.

The interface supports yes-or-no questions, choices among named options and ratings on an ordered scale. That makes the output suited to a software workflow’s next step: which team gets a ticket, whether a message should be flagged, or whether a request should be approved. Several questions can share the same input.

Base Labs built joey-1 by further training Qwen3.5-2B, not by starting from scratch. The model card describes 6.21 million decision examples and training that updated the language model while leaving its image-processing component unchanged. Safety and abstention made up 16.1% of the mixture, and training included an “I don’t know” option for uncertain answers.

The economics depend on volume—and uncertainty

Baseten’s cost example is a support operation classifying 10 million tickets a month, each around 500 tokens—the units of content a model processes. That totals 5 billion tokens. The company puts the model bill at about $50 a month on joey-1, compared with $500 on its 27-billion-parameter mammoth-1.

Baseten recommends joey-1 as a first-line model, with elephant-1 or mammoth-1 handling cases where it is uncertain. The company also targets decisions that must finish before another operation can begin, because that waiting time is added to every response. Its reported latency comparison illustrates the intended benefit.

Speed is not the same as capability. Joey-1’s model card reports a score of 63.30 on the Jev Decision Index 0.4 public split, placing it 37th among 230 models. It claims the highest score among models under 4 billion parameters on that split—not the highest score overall.

The card’s comparison puts elephant-1 at 72.24, above joey-1 in every evaluated area. That gap supports the family’s stated division of labor: joey-1 prioritizes inexpensive, fast calls, while the larger sibling provides a more capable option. These are published benchmark results, not a guarantee about a particular customer’s decisions.

Two hosted routes, plus open weights

• Model APIs: The shared service supports a 16,000-token context window, the amount of material handled in one request. Baseten says zero data retention is enabled by default.
• Dedicated deployments: Baseten advertises the full 262,000-token context window and support for customized weights. It says workload-specific tuning and placing models near related workloads can improve serving speed further.

Developers can also download the MIT-licensed weights from Hugging Face. Joey-1 and elephant-1 are available for further training through Baseten Training Jobs, with a published fine-tuning cookbook. All three family members share an API; Baseten says moving between them requires a one-line change.