Innovation & Tech

Baseten's Bison-1 Tops Decision Index With 0.14 Brier Score

Baseten's bison-1 leads Artificial Analysis's Decision Index v2.1 with a 0.14 Brier score. Both new models offer 128K context, image input and calibrated probabilities from $0.035 per million tokens.

By Amara Osei

3 min read

Baseten Extends Decision Models to 128K Context and Image Input
Baseten Extends Decision Models to 128K Context and Image InputAI-generated

What's News

  • Baseten released bison-1 and groundhog-1 on November 20, 2026, post-trained from Google's Gemma 4.
  • Bison-1 scored 76 on Artificial Analysis Decision Index v2.1, first among 19 models, with the lowest Brier score of 0.14.
  • Both models support 131,072 tokens — eight times the 16K context of earlier mammoth-1, elephant-1 and joey-1 APIs.
  • Pricing is $0.10 per million tokens for bison-1 and $0.035 for groundhog-1.
  • Training used 6.71 million decision examples on 192 B200 GPUs for bison-1 and 128 for groundhog-1.

Baseten's bison-1 scored 76 on Artificial Analysis's Decision Index v2.1, taking first place among 19 evaluated decision models with the index's lowest Brier score at 0.14. The company released bison-1 and groundhog-1 on November 20, 2026 — two open-weight models post-trained from Google's Gemma 4 that return calibrated probabilities for predefined decisions instead of free-form text.

The timing matters. Artificial Analysis published its results on November 19, one day before the release. D1.2-ultra came second at 75 and Baseten's own mammoth-1 scored 71. The Brier measure calculates the mean squared gap between the probability a model assigns to its chosen answer and whether that answer is actually correct — bison-1's 0.14 beat d1.2-ultra's 0.15, meaning its confidence tracks outcomes more closely.

What are the two models, exactly?

Bison-1 is a dense 31B-parameter model post-trained from Gemma 4 31B. Groundhog-1 is a 26B mixture-of-experts model based on Gemma 4 26B-A4B, running about 4B active parameters per token. Baseten prices them at $0.10 and $0.035 per million tokens respectively, matching the older mammoth-1 and elephant-1 tiers.

Both accept text and images and support 131,072 tokens — eight times the 16K context of Baseten's earlier mammoth-1, elephant-1 and joey-1 APIs. A single request bundles one state with any number of yes/no, named-choice or ordered-score questions into one shared-prefix prefill, with no text generation required.

They're available through Baseten Model APIs and as weights on Hugging Face. Licensing is split: Google's Gemma 4 base models carry Apache License 2.0, while Baseten's modifications ship under MIT, with use subject to both.

How does the Decision Index work?

Artificial Analysis's Decision Index v2.1 combines six independent evaluations. The breakdown:

  • AA-Paperclip — decisions and extraction across 1,500 resumes, receipts, invoices and research papers — carries 30% of the weight
  • When2Call, a tool-calling evaluation, carries 10%
  • Workplace judgement, knowledge and hallucination, reasoning, and decision invariance each carry 15%

The Decision Score subtracts a Brier calibration penalty from accuracy. Because the underlying probabilities support confidence thresholds — answering some cases and routing others elsewhere — calibration directly shapes how a decision system gets deployed.

Is bison-1's lead uniform?

No. The model card reports first place on When2Call, EvalSafe-O*NET-AA, HLE-MultiChoice v1.1 and AA-Symmetry v1.1, plus second on AA-Paperclip. But its 60 on AA-Omniscience-Decision v1.1 ranked 11th. The aggregate lead includes substantial variation across components, not uniform dominance.

Groundhog-1's card states it has not yet received a Decision Index score at all, leaving its mixture-of-experts performance unverified.

What went into training?

Baseten trained both models 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 includes a trained "I don't know" option. About one example in seven carries a probability target rather than a single right answer.

Baseten attributes a significant calibration gain to expanding these soft-label targets alongside the Gemma base. The compute bill: 192 B200 GPUs for bison-1 and 128 for groundhog-1.

On Baseten's own long-context runs at 128K tokens, bison-1 scored 84.8 on LongMemEval M and 87.1 on LongMemEval S. Groundhog-1 scored 77.4 and 79.3 respectively. These company-run tests sit separate from Artificial Analysis's six-part index.

What happens next?

The release pushes long-context, multimodal decision making across both of Base Labs' established price tiers. One engineering choice stands out for production users: requests above the trained length limits get rejected rather than truncated, preserving calibrated outputs.

Groundhog-1's missing independent index score is the open question — it will determine whether the mixture-of-experts design delivers the mammoth-1-level performance Baseten expects.

Source: unite.ai

Share this article:

More from Amara Osei

Amara Osei

Show full bio

Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.

2,503 articles

Related articles

« Previous article