Baseten Releases joey-1, a 2B Decision Model at $0.01 per 1M Tokens
Oct 29, 2026, 2:20 PM · MarkTechPost
Baseten's joey-1 skips the chatbot act and just answers the question with a label and a confidence score, which is exactly what most production AI calls should look like.
Why it matters
Baseten's research lab, Base Labs, released joey-1, a 2B-parameter decision model post-trained from Qwen3.5-2B. Instead of writing text, it returns a probability for every option on typed questions: pick one, yes or no, or an ordered score such as urgency. It is priced at $0.01 per million input tokens on Baseten's Model APIs, with open weights on Hugging Face under MIT.
Baseten reports a 20 millisecond median latency on a typical 2,000-token request and a score of 63.30 on the public split of the Jev Decision Index 0.4, which it says is the top result for any model under 4B. These are self-reported numbers from Baseten's model card.
From the desk
We like this design, and we think it is overdue. A huge share of real AI traffic is not conversation. It is routing a ticket, flagging a message, deciding whether something needs a human. Using a chat model for that means paying for prose and then parsing it. A model that answers in one forward pass with a calibrated probability is cheaper, faster, and easier to audit.
Calibration is the key feature, not a footnote. Baseten trained it with a calibration term and an explicit "I don't know" option, so a team can set a confidence threshold and send uncertain cases up to its larger siblings, elephant-1 or mammoth-1, or to a person. That is the right shape for an AI system: let the cheap model handle the obvious, escalate the rest.
The price makes the case concrete. Baseten's own example puts ten million support tickets a month at about $50 on joey-1. At that cost, nobody will think twice before classifying everything.
That is also the downside. When a decision costs almost nothing, it gets made about everything and everyone, often without a human seeing it. A guardrail check that wrongly blocks a user, a router that quietly deprioritizes certain requests, a score that decides urgency. Errors in a 2B model are cheap per call and costly in aggregate. And joey-1 trails the 9B elephant-1 by about nine points on the same index, so it really is a first-line filter, not a final judge.
Where this leads if it scales: decision models become the invisible plumbing of most AI products. I'm watching for whether teams actually use the confidence scores to escalate, or just take the top label and move on. The tool supports doing this well. Whether people do is a different question.
Context
joey-1 is the third model in a family that includes mammoth-1 (27B) and elephant-1 (9B), released the week before. All three share training data and the same API, so swapping between them is a one-line change. Baseten says it was trained on 6.21 million decision examples on two nodes of eight H100s, and it can be self-hosted on a single CUDA GPU. It is also listed on OpenRouter.
Who feels it
- Developers
- Typed questions with probabilities remove output parsing and make thresholds and escalation simple to implement.
- Enterprises
- High-volume classification and guardrail checks get dramatically cheaper, but audit logs and human escalation paths still matter.
- End users
- More automated decisions will touch their requests unseen; wrong calls from a small model can be hard to notice or appeal.
What to watch
- Independent latency and accuracy tests outside Baseten's own platform
- Whether competing decision models from Perplexity, Cloudflare, and others respond on price
- Fine-tuned joey-1 variants appearing for specific domains
- How teams report using the confidence scores and the "I don't know" option in production