Most model launches compete on the quality of generated text. Baseten’s newest pair competes on something narrower and, for a lot of production systems, more useful: the quality of a decision.
bison-1 and groundhog-1 are decision models. They do not write paragraphs or answer in sentences. A caller sends a state, text, optionally with images, and a set of questions. The model returns a probability for each answer in a single pass, without generating output tokens. The questions can be yes/no, a choice among named options, or a score on an ordered scale.
That makes them a different product from a chat model. A support system can ask whether a ticket is a refund request, which team should handle it and how urgent it is, all in one call. A guardrail can ask whether a prompt is a jailbreak attempt. A router can ask which model should handle a request. The output is typed and ready for code, not prose that another parser has to interpret.
What Baseten released
bison-1 is a 31B dense model post-trained from Gemma 4 31B. groundhog-1 is a 26B mixture-of-experts model with about 4B active parameters, built on Gemma 4 26B-A4B. Both accept text and images, both handle up to 131,072 tokens, and both are available through Baseten’s Model APIs with open BF16 weights on Hugging Face.
The context window is eight times larger than Baseten’s previous decision models, mammoth-1 and elephant-1. That matters for the kind of work decision models are asked to do. The detail that changes a decision is often buried far from the question: a canceled order in an earlier email, an exception in a policy document, a prior incident in a long history. A model that only sees a summary can make a confident wrong call.
The trade-off is size. Baseten lists no quantized builds or inference GPUs for the open weights. By the company’s own arithmetic, the BF16 weights are about 62 GB for bison-1 and 52 GB for groundhog-1. These are not laptop models. The API is the practical route for most teams.
The benchmark claim
On Artificial Analysis’s Decision Index v2.1, bison-1 scores 76, first among 19 models. Liquid AI’s d1.2-ultra scores 75, and Baseten’s own mammoth-1 scores 71. The index blends six evaluations, with AA-Paperclip carrying 30% of the weight.
The component scores show where the model is strong and where it is not. bison-1 ranks first on When2Call (98), EvalSafe-O*NET-AA (93), AA-Symmetry v1.1 (89) and HLE-MultiChoice v1.1 (51). It ranks second on AA-Paperclip (73). Its weakest result is AA-Omniscience-Decision v1.1, where it scores 60 and ranks 11th. Knowledge-heavy decisions are the soft spot.
Calibration is the more interesting number. bison-1 has a Brier score of 0.14, the lowest on the index, ahead of d1.2-ultra at 0.15. Lower is better: the score measures how closely stated confidence tracks actual accuracy. A model that says 80% and is right 80% of the time is more useful in an automated pipeline than a model that says 99% and is right 70% of the time.
Baseten credits two things for the calibration. One is the Gemma base. The other is a heavier use of soft labels in training: instead of teaching the model that one answer is always correct, about one example in seven carries a probability target. The training objective also includes a Brier calibration term and a trained “I don’t know” option. The result is a model that can express uncertainty instead of forcing a confident guess.
Price and positioning
bison-1 costs $0.10 per million input tokens. groundhog-1 costs $0.035. Both use the same API as Baseten’s older decision models, so switching is a model-name change. The IDs are decision-models/baseten-bison-latest and decision-models/baseten-groundhog-latest.
The pricing puts bison-1 in the same bracket as mammoth-1, and groundhog-1 in the same bracket as elephant-1. The choice is less about a single ranking and more about the job. bison-1 is for accuracy and long context. groundhog-1 is for lower latency and cost, with long context as the main upgrade over the older small model.
There is also a provenance argument. Baseten’s previous models sat on Qwen backbones. The new pair is the company’s first on what it calls an American backbone, and the launch post frames that as a deliberate choice for teams that want a Western base model. Model provenance is becoming a procurement question, not just a technical one.
What the numbers do not settle
groundhog-1 has no official Decision Index score yet. Baseten expects it to match mammoth-1, but that is an expectation, not a measurement. The benchmark numbers for both models are largely self-reported, even when an outside index is involved. A single index is a useful filter, not a deployment decision.
There is also a practical detail for anyone building on the API: requests longer than the trained limit are refused, not truncated. Pipelines that assume silent truncation will need a length check. That is a small engineering note with a large failure mode.
The bigger question is whether decision models become a standard layer in production stacks. The case is straightforward. Many systems do not need another writer; they need a cheap, fast, well-calibrated classifier that can read a long context and return a structured answer. bison-1 is the strongest public claim yet that such a model can be both accurate and honest about its confidence. The caveats are real, but the category is now easier to take seriously.
Sources: Baseten; Base Labs; bison-1 model card; groundhog-1 model card; MarkTechPost.
Have a tip? Email [email protected] or use the contact form and choose Editorial. We do not promise anonymity.
Read next: How AI benchmarks actually work, and why a 0.46-point lead is not a verdict; Google opens Gemini 4 Carbon to all paying developers; Claude picked a bombed Iranian school as a target in tests; SlipStream: making high-bandwidth flash work for AI chips. Browse all Artificial Intelligence coverage.
