LATEST
AI Tools

Baseten Bison-1 Tops the Decision Index at $0.10 per Million Tokens

Updated November 21, 2026
13 min read
Editor's Note: This article was human-reviewed for accuracy
A man studying a chess board, weighing his next move

Quick Answer & Key Takeaways

  • • Baseten's bison-1 is the top-ranked decision model on the Artificial Analysis Decision Index v2.1, scoring 76 out of 19 models evaluated.
  • • Both new models take 128,000 tokens of text and images, eight times the context of Baseten's earlier decision models.
  • • Bison-1 costs $0.10 per million tokens and groundhog-1 costs $0.035, with open weights on Hugging Face.

Baseten released two new decision models on November 20, 2026, and the bigger one arrived with a headline worth checking: bison-1 scored 76 on the Artificial Analysis Decision Index v2.1, the top score among the 19 decision models evaluated. Its smaller sibling, groundhog-1, costs about a third as much and takes the same 128,000-token context. Both are live on Baseten Model APIs, and both ship with open weights on Hugging Face.

That is a narrow claim, and the narrowness is the point. These models do not write essays, hold conversations, or generate code. They answer predefined questions with calibrated probabilities, which makes them useful for the kind of work that usually gets handed to a full chat model at ten times the price: triage, routing, moderation, and any pipeline step where the answer is a choice rather than a paragraph. The question for teams is not whether bison-1 is a good model in general. It is whether a decision-only model belongs in their stack, and at what volume the per-token price starts to matter.

What a decision model actually does

A decision model takes a state and a set of questions, then returns a probability for every option. The state can be text, optionally with images. The questions can be yes-or-no, a choice among named options, or an ordered score such as an urgency level. All questions in a request are answered from one shared-prefix prefill, and the model generates no text at all. That last detail is what keeps the cost down: there is no output token bill, because there are no output tokens.

Baseten's own example in the model card makes the shape concrete. A support team sends a customer message as the state, then asks three questions at once: is this a refund request, which team should handle it, and how urgent is it. One call returns a probability for each answer. The same pattern covers ticket triage, spam detection, jailbreak guardrails, and routing between models, which Baseten listed as target tasks when it launched the earlier generation of these models.

Calibration is the feature Baseten leans on hardest. A model that says it is 83 percent sure should be right about 83 percent of the time, and Baseten trained these models with soft probability targets, about one example in seven carrying a probability instead of a single right answer, plus a trained "I don't know" option. That matters when the output feeds an automated threshold: a confident wrong answer auto-approves a payment, while an honestly uncertain one gets routed to a person.

The two new models at a glance

Bison-1 is a dense 31-billion-parameter model post-trained from Google's Gemma 4 31B. Groundhog-1 is a 26-billion-parameter mixture-of-experts model built on Gemma 4 26B-A4B, with about 4 billion active parameters per token. Both accept text and image input, both run to 131,072 tokens of context, and both use the same API as Baseten's earlier decision models, so switching from mammoth-1 or elephant-1 means changing one model name.

ModelParametersBase modelContextPrice per 1M tokens
bison-131B denseGemma 4 31B128K$0.10
groundhog-126B MoE, ~4B activeGemma 4 26B-A4B128K$0.035
mammoth-127BQwen16K$0.10
elephant-19BQwen16K$0.035
joey-12BQwen3.5-2B16K$0.01

The context jump is the clearest upgrade over the earlier models. Sixteen thousand tokens forced teams to select excerpts or summarize long documents before asking a question. At 128,000 tokens, a full policy document or a long conversation history fits in one request. Baseten's own example is an invoice that looks ready for payment until an earlier email shows the order was canceled, a detail that only surfaces if the whole history is in context. Both models were trained on sequences up to 131,072 tokens, and requests longer than the trained limits are refused rather than silently truncated, which is a behavior worth knowing before you build around it.

There is also a framing choice in the launch worth naming plainly. Baseten calls Gemma 4 an "American backbone" and positions the pair as a way for teams to choose American models on merit. Gemma 4 is Google's open model, and the earlier mammoth-1, elephant-1, and joey-1 were built on Qwen backbones. The "American" label is Baseten's positioning, not a technical property of the weights. What matters for buyers is the performance and the price, and those are measurable.

What the benchmark says, and what it does not

The headline number comes from Artificial Analysis, an independent benchmarking organization. Its Decision Index v2.1 combines six evaluations: document decisions across resumes, receipts, invoices, and research papers; tool-calling decisions; workplace judgement; knowledge and hallucination; reasoning; and decision invariance. The index subtracts a Brier calibration penalty from accuracy, so a model that is confidently wrong scores worse than one that is equally accurate but honest about its uncertainty. Bison-1 scored 76, ahead of Liquid AI's d1.2-ultra at 75 and Baseten's own mammoth-1 at 71, and its Brier score of 0.14 was the lowest on the index, against 0.15 for d1.2-ultra.

Benchmark componentBison-1 scoreRank of 19
When2Call (tool calling)981
EvalSafe-O*NET-AA (workplace judgement)931
AA-Paperclip (document decisions)732
AA-Symmetry v1.1 (decision invariance)891
HLE-MultiChoice v1.1511
AA-Omniscience-Decision v1.16011

The component table is where the honest reading lives. Bison-1 leads or nearly leads on five of the six components, but it ranks 11th of 19 on AA-Omniscience-Decision, the knowledge-heavy component, with a score of 60. The aggregate lead is real, but it is not uniform dominance, and a team whose decisions hinge on broad factual knowledge should weigh that component more than the headline.

Two more caveats belong in the same paragraph as the score. First, groundhog-1 has not been scored on the Artificial Analysis Decision Index yet, so its cheaper price comes without an independent number. Baseten says it expects groundhog-1 to match mammoth-1 on the index, which would put it around 71, but that is the vendor's expectation, not a measured result. Second, the long-context numbers are Baseten's own runs, not independent ones: at 128,000 tokens, bison-1 scored 84.8 on LongMemEval M and 87.1 on LongMemEval S, while groundhog-1 scored 77.4 and 79.3. Treat those as company-reported until an independent lab repeats them.

For scale, the same index family puts other decision models close behind: Nace's Drex Max v1.1 scored 73, and Liquid AI's d1.2-ultra sits at 75. The gap between first and third is five points, which is a real lead but not a different league. Our earlier review of Drex Max v1 covers a free-weight alternative in the same category, and our Liquid AI LFM3 review covers that lab's open-weight family.

Pricing: what $0.10 and $0.035 actually buy

Pricing is per million input tokens, with no output charge because the models produce no output tokens. Bison-1 costs $0.10 per million tokens, the same as mammoth-1. Groundhog-1 costs $0.035, the same as elephant-1. The launch post lists no free allowance, so billing is per token from the first request, and teams should confirm the current rate on Baseten's pricing page before committing volume.

ModelUSD per 1M tokensGBP per 1M tokens*CAD per 1M tokens*
bison-1$0.10£0.08C$0.14
groundhog-1$0.035£0.03C$0.05

*Approximate conversions at November 20, 2026 market rates (1 USD = 0.762 GBP, 1.43 CAD), verified at publication.

The arithmetic that matters is per decision, not per million tokens. A million decisions with a 1,000-token state each would use a billion tokens, which costs about $100 on bison-1 and about $35 on groundhog-1, by our calculation from the published rates. A million decisions with a 10,000-token state each costs ten times that. Because the price is per token rather than per decision, long states are where the bill grows, and the 128K context window is a double-edged feature: it lets you include everything, and it charges you for everything you include.

There is also a cheaper tier in the same family that teams should not overlook. Joey-1, at $0.01 per million tokens, handles the smallest decisions for a tenth of bison-1's price, and it keeps the same API. The right question is not which model costs the least in general, but which model is accurate enough for the decision at hand. A spam check that joey-1 answers correctly does not need bison-1, and a payment approval probably does not belong on joey-1.

How to call them

Both models are live on Baseten Model APIs under the IDs decision-models/baseten-bison-latest and decision-models/baseten-groundhog-latest. They use the same API as mammoth-1 and elephant-1, so existing integrations change one model name. The weights are also open on Hugging Face, under the baselabs organization, for teams that would rather self-host.

A request carries one state and any number of questions. Each question declares a type: yes-or-no, a choice among named options with a short description for each, or an ordered score with levels listed lowest first. The response returns the chosen answer and the probability for every option. Requests longer than the trained length limits are refused, not truncated, so a pipeline that occasionally sends oversized states needs to handle a rejection rather than assume a partial answer.

Training details, published on the model cards, give a sense of what went into the pair. Both were trained on 6.71 million decision examples covering core decision skills, safety and abstention, judging and expert domains, documents, and long-context tasks, including public benchmark sets. The objective combines cross-entropy with a Brier calibration term, and about one example in seven carries a probability target instead of a single right answer. Bison-1 trained on 192 B200 GPUs; groundhog-1 on 128.

Licences and the open-weights fine print

The weights are open, but the licence stack has two layers, and teams should read both. Gemma 4, the base model, is licensed under Apache License 2.0. Baseten's modifications are released under the MIT License. Use of the models is subject to both licences. The model cards also state that Gemma is a trademark of Google LLC, and that the models are not affiliated with or endorsed by Google.

That is a permissive combination for most commercial use, and it compares well with decision models shipped under revenue-capped or non-commercial licences. The practical check is still the same as with any open-weight release: confirm the licence terms fit your deployment, keep the attribution intact, and remember that open weights do not mean a managed service. Self-hosting means running the model yourself, while the hosted API means paying Baseten per token and accepting its data handling terms.

Where they fit, and where they do not

Decision models earn their keep in pipelines where the same judgement repeats thousands of times a day. Ticket triage, content moderation queues, fraud signals, document classification, and routing between larger models are the natural fits. In each case the output is a choice with a confidence number, and the confidence number is what lets a team auto-handle the clear cases and escalate the uncertain ones.

They do not replace a general-purpose chat model. If the task needs generated text, a summary, a draft reply, or an open-ended answer, a decision model has nothing to offer, because it returns probabilities, not words. They also do not replace a human on high-stakes calls. A calibrated 83 percent is still a 17 percent chance of being wrong, and for decisions where a wrong answer is expensive, the right design is a threshold that sends the uncertain cases to a person, not a model that answers everything.

Against the wider decision-model field, bison-1's position is the top of a crowded board rather than a runaway lead. Liquid AI's d1.2-ultra trails by one point, Nace's Drex Max v1.1 by three, and Baseten's own mammoth-1 by five. The meaningful differences are the 128K context, which none of the earlier Baseten models had, and the image input, which opens document decisions that include scans and photos. Groundhog-1's case is thinner until it gets an independent score, but its price matches elephant-1 while offering eight times the context, so it is the more interesting of the two for teams already running elephant-1 at volume.

Common mistakes before you switch

First, do not treat the headline score as a guarantee for your data. The index is six evaluations; your decisions are a seventh, unmeasured one. Run your own labelled sample through both models before moving a live pipeline, and compare the calibration on your cases, not just the accuracy.

Second, do not fill the 128K window by default. Every token in the state is billed, and a long context that adds no signal is a cost with no benefit. Measure accuracy with a short state and a long state on the same sample, then include only what moves the result.

Third, do not assume groundhog-1 matches mammoth-1 because Baseten expects it to. The expectation is plausible, the mixture-of-experts design is cheaper to run, and the price is identical to elephant-1, but until Artificial Analysis publishes a score, treat groundhog-1 as unranked.

Fourth, plan for refusals. Overlong requests are rejected outright, so a state that occasionally exceeds the trained length needs a trimming step and a fallback, not a retry loop that keeps sending the same oversized payload.

Is bison-1 really the best decision model?

On the Artificial Analysis Decision Index v2.1, yes, by the measured score: 76, first of 19 models, with the lowest Brier calibration score on the index at 0.14. That is an independent ranking, but it covers six evaluations, and bison-1 ranks 11th on the knowledge-heavy component. "Best" here means best on this index, not best for every decision.

How much does it cost to run a decision through bison-1?

Bison-1 costs $0.10 per million input tokens, with no output charge. A 1,000-token decision costs about $0.0001, and a 10,000-token decision about $0.001, by our calculation from the published rate. Groundhog-1 costs $0.035 per million tokens, so the same decisions cost about a third as much.

What is the difference between bison-1 and groundhog-1?

Bison-1 is a dense 31-billion-parameter model and the more accurate of the two, with the top score on the Artificial Analysis Decision Index. Groundhog-1 is a 26-billion-parameter mixture-of-experts model with about 4 billion active parameters per token, priced at $0.035 per million tokens, and it has not been scored on the index yet. Both take 128,000 tokens of text and images.

Can I run these models on my own hardware?

Yes. The weights are open on Hugging Face under the baselabs organization. The Gemma 4 base is under Apache License 2.0 and Baseten's modifications are under the MIT License, so use is subject to both. Self-hosting means running the model yourself; the hosted API is the pay-per-token alternative.

Do decision models generate text?

No. They return a calibrated probability for each option in a predefined question, without generating text. That is why there is no output token charge, and also why they cannot draft replies, summarize documents, or hold a conversation. They answer choices; a chat model still does the writing.

Should I switch from mammoth-1 or elephant-1?

If your decisions need more than 16,000 tokens of context or image input, yes, the switch is straightforward, since both new models use the same API and the same prices as the older pair. If your current setup works within 16K and you have no image inputs, the upgrade buys little beyond the benchmark gains, and a side-by-side test on your own data is the safer first step.

Baseten's bison-1 earns its headline on a measured, independent index, and groundhog-1 extends the same long-context design to a much lower price. The pair is a genuine option for teams that make the same judgement thousands of times a day and want a number they can threshold instead of a paragraph they have to parse. Run your own labelled sample first, keep the context short enough to justify its cost, and let the calibration, not the ranking, decide whether it stays in your pipeline. You can read the launch post on Baseten's blog and the model cards for bison-1 and groundhog-1 on Hugging Face.

Sources

Exchange rates: November 20, 2026 market rates, verified at publication. Prices are Baseten's published per-token rates; confirm the current rate on Baseten's pricing page before committing volume.

13 min read
27 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more