Quick verdict
Verdict up front: bison-1 and groundhog-1, released by Baseten on November 20, 2026, are not chatbots — they are open-weight decision models that return calibrated probabilities for yes/no, choice and scoring questions, the unglamorous work behind triage, routing and approval pipelines. bison-1 ranks #1 of 19 on the Artificial Analysis Decision Index v2.1 and both models jump to a 128k context window with image input, at $0.10 and $0.035 per million tokens. If your app asks a model to decide rather than write, this is currently the pair to beat — with the caveat that groundhog-1 has no official index score yet.
What are decision models, and what is new here?
A decision model answers one fixed question with a calibrated probability: is this ticket urgent, is this message spam, should this payment be approved, should this request go to a human. Baseten’s Base Labs family post-trains base models specifically for that job. The new pair switches the backbone from Qwen to Google’s Gemma 4 — bison-1 is a 31B dense model built on Gemma 4 31B, groundhog-1 a 26B mixture-of-experts on Gemma 4 26B-A4B with about 4B active parameters — a deliberate move to an American base model. Both are live on Baseten Model APIs using the same API as the older models (switching is a model-name change), and both sets of weights are open on Hugging Face.
The practical upgrades over the previous mammoth-1 and elephant-1 are context and inputs: 128k tokens instead of 16k — eight times more, enough for a full policy or conversation history instead of excerpts — and image input alongside text. Baseten’s example is telling: an invoice may look payable until an email buried in the history shows the order was canceled.
What the pricing doesn’t make obvious
Verified on Baseten’s own launch post: the $0.10 and $0.035 figures are per million input tokens, and decision models produce no output tokens — there is no second, output-side charge hiding behind the headline price, unlike the chat models most teams are used to budgeting for. Pricing is also deliberately unchanged from the previous generation (bison-1 costs what mammoth-1 cost; groundhog-1 what elephant-1 cost), so the 8x context jump is effectively free for existing users. One honest gap: Baseten says it expects groundhog-1 to match mammoth-1 on the official index — it has no official Decision Index score yet, so its ranking is an estimate, not a measurement.
bison-1 vs groundhog-1: the numbers
| bison-1 | groundhog-1 | |
|---|---|---|
| Parameters | 31B (dense) | 26B MoE, ~4B active |
| Backbone | Gemma 4 31B | Gemma 4 26B-A4B |
| Context window | 128k | 128k |
| Inputs | Text and images | Text and images |
| Price per 1M tokens | $0.10 | $0.035 |
| Decision Index v2.1 | 76 — #1 of 19 | No official score yet |
| Use it when | Highest accuracy, long context | Long context at low latency and cost |
On calibration — how close the model’s stated confidence is to how often it is right — bison-1 posts the index’s best Brier score at 0.14, ahead of d1.2-ultra at 0.15. It is the same specialized-model logic as Cloudflare’s extraction models in our Clairic-flash review: a small model trained for one narrow job can beat a generalist at it, for a fraction of the price.
Strengths
- Top-ranked accuracy. A Decision Index score of 76 puts bison-1 first among 19 decision models, one point ahead of d1.2-ultra (75) and five ahead of mammoth-1 (71).
- Best-in-index calibration. For decisions that trigger real actions — approvals, escalations, blocks — a probability you can trust matters more than raw accuracy.
- 128k context with images. Full histories and documents in one call, at the same price as the 16k generation.
- Open weights, same API. Weights are on Hugging Face for self-hosting (browse the ecosystem via our Hugging Face review), and hosted users switch by changing one model name.
- Input-only pricing. No output tokens means simple, predictable per-decision costs.
Weaknesses
- groundhog-1 is unproven on the index. Baseten expects it to match mammoth-1, but there is no official score yet — evaluate it on your own decisions before trusting it in production.
- Vendor-reported numbers. The headline scores come from Baseten’s launch post citing Artificial Analysis; independent replication is still thin a day after release.
- Narrow by design. These models do not write, chat or explain — if you need a rationale in prose, you still need a general model beside them.
- Self-hosting is heavy. A 31B dense model is a real GPU commitment; the hosted API is the realistic route for most teams.
Who are bison-1 and groundhog-1 for?
Product and platform teams running classification, triage, routing, moderation or approval decisions at volume — support queues, fraud checks, content pipelines — where a calibrated probability plugs straight into business logic. Pick bison-1 when the decision is expensive to get wrong; pick groundhog-1 when latency and cost dominate and the decision is cheap to reverse. General chatbot builders should look elsewhere; this is a specialist tool.
FAQ
What is a decision model?
A model trained to return a calibrated probability for a fixed question — yes/no, a choice, or a score — instead of generating open-ended text. Baseten’s models are built for triage, routing and approval-style decisions.
How much do bison-1 and groundhog-1 cost?
bison-1 costs $0.10 per million tokens and groundhog-1 $0.035 per million tokens on Baseten Model APIs. Decision models produce no output tokens, so the price is effectively input-only. Open weights on Hugging Face are free to download and self-host.
Are bison-1 and groundhog-1 open source?
They are open-weight: both models’ weights are published on Hugging Face, and both are also available hosted on Baseten Model APIs using the same API as Baseten’s earlier decision models.
What context window do they support?
Both support 128k tokens on Model APIs — eight times the 16k of the previous mammoth-1 and elephant-1 — and both accept text and image input.
Bottom line: bison-1 is the new decision-model benchmark leader with the index’s best calibration, and groundhog-1 offers the same 128k, multimodal package at roughly a third of the price. For teams whose models decide more than they write, this pair — launched November 20 — should be the first evaluation on the list.

