Baseten's decision model lineup got an upgrade on November 20. bison-1, a 31-billion-parameter model built on Google's Gemma 4 architecture, now sits at the top of the Artificial Analysis Decision Index v2.1 with a score of 76, edging out the previous leader d1.2-ultra at 75 and besting every other specialized decision model that has shipped on the benchmark. At $0.10 per million input tokens, it matches OpenAI's Decisions API price exactly, and it ships as open weights that any team can download, fine-tune, or self-host without negotiating a contract.
bison-1 and groundhog-1: Bigger Context, Gemma Backbone, Open Weights
Baseten's research arm, Base Labs, launched two models in the same November 20 announcement. bison-1 is the flagship: a 31B-parameter model fine-tuned from Gemma 4 31B, with a 128k-token context window and pricing at $0.10 per million tokens on Baseten Model APIs. groundhog-1 is the efficiency option: a 26B mixture-of-experts model with roughly 4 billion active parameters, the same 128k context, and pricing at $0.035 per million tokens. Both accept text and image input, both use the same API endpoint as Baseten's existing decision models, and both have published weights on Hugging Face under the Base Labs organization, available for self-hosting or fine-tuning. The Base Labs announcement post confirms the release date and all model specifications cited here.
The 128k context is the number Baseten leads with, and for good reason. Their prior generation, mammoth-1 and elephant-1, capped at 16k tokens per request, as described in the October launch post. The 8x expansion matters because agent decisions rarely live in isolation. An approval pipeline might need a 60-page vendor contract to decide whether a purchase order is authorized. A content moderation system needs a user's message history, not just the current post, to judge whether a borderline message crosses a line. A customer support router needs the full CRM thread to know whether this escalation should go to billing or to retention. At 16k, teams had to summarize or chunk all of that history before passing it to the decision model, adding latency and introducing summarization errors that degraded accuracy in ways that were hard to predict. At 128k, the full document or thread fits in a single call.
The backbone switch from Qwen to Gemma 4 is also deliberate. mammoth-1 and elephant-1 were post-trained from Qwen 3.5 at 27B and 9B respectively. Baseten's move to Gemma 4 reflects what the company frames as a commitment to American-designed base models, responding to enterprise procurement teams that increasingly track the provenance of AI components in their stacks. The post states that bison-1 and groundhog-1 are a starting point for more research on American backbones, suggesting additional Gemma-family releases are planned. For teams that already deploy on Baseten's Model APIs, the migration path is a model name change, nothing more.
How Soft Labels Made bison-1 the Most Calibrated Decision Model
Decision models work differently from language models. They do not generate text. A single request can carry multiple questions and returns probabilities, yes-or-no scores, or ranked choices without any generated prose. This design is why they can process a large number of questions in a single call and return results in roughly 150 milliseconds, far faster than prompting a full language model and parsing its response. The tradeoff is that they can only handle bounded decisions, choices over a defined option set, rather than open-ended generation tasks.
The key advance in bison-1 is calibration. The Artificial Analysis Decision Index v2, which launched November 8 and defines its headline metric as 50 times the quantity (1 plus accuracy minus Brier score), rewards both getting the right answer and knowing how confident to be. A lower Brier score means stated confidence is closer to how often the model is correct. bison-1 achieves a Brier score of 0.14, compared to 0.15 for the previous leader d1.2-ultra. That gap sounds small, but across millions of agent decisions in production, a systematic miscalibration of one point in Brier score compounds into real accuracy losses in the downstream actions that depend on those probabilities.
Baseten traces the calibration gains to what it calls soft-label training. Rather than training on a single correct answer per question, the training data records probability distributions over options: 83% answer A, 17% answer B, based on how many human or model raters agreed. This teaches the model that humans disagree, and it builds that uncertainty into the weights rather than forcing artificial confidence in every answer. mammoth-1, launched in October with a score of 71 on the AA index, was the previous Base Labs benchmark peak. bison-1 at 76 represents a five-point improvement and leapfrogs every model that entered the index between October and today, including the previous leader at 75. The index currently covers 19 decision models. Baseten's post is the primary source for the v2.1 rankings, so verification against the live Artificial Analysis leaderboard is advisable before treating the score as independently audited.
OpenAI Charges $0.10 Per Million for GPT-6 Luna. So Does Baseten.
OpenAI made its Decisions API generally available on October 22, 2026, at $0.10 per million input tokens, powered by GPT-6 Luna. The OpenAI developer changelog confirms the beta opened October 6 after a limited preview at DevDay on September 29, and the GA announcement coverage details the full feature set that shipped with general availability: prompt caching, image inputs via JSON, support for up to 400 questions per request, Zero Data Retention for HIPAA-eligible workloads, and native integration with GPT-Live for real-time decision delegation inside voice sessions. Response latencies ran around 150 milliseconds in demonstrations, and OpenAI claims the endpoint is roughly 10 times faster than the Responses API for bounded decision tasks.
bison-1 sits at the same $0.10 price. It does not have native integration with ChatGPT Wallet or GPT-Live, features OpenAI wired into its proprietary stack between beta and GA. What bison-1 has is an API-equivalent setup on the hosted side, open weights for teams with infrastructure to self-host, and a benchmark result that now matches or exceeds the proprietary alternative on the leading decision-model index. groundhog-1 adds a second data point: at $0.035 per million tokens, it is 65% cheaper than OpenAI's entry price and 65% cheaper than bison-1, targeting the high-volume routing and classification tier where even small per-call cost differences compound into five-figure annual cost differences at the routing volumes most enterprise stacks process.
The practical comparison for a team choosing today: a GPT-6 Luna Decisions API call costs $0.10/M and runs on a proprietary model they cannot inspect, fine-tune, or move off-platform without rebuilding the integration from scratch. A bison-1 call on Baseten's hosted API costs the same $0.10/M, using weights the team can download and fine-tune on internal data, or self-host on their own compute if usage volume makes that economical. For regulated industries like healthcare and financial services, the ability to audit the model weights and control where inference happens is worth more than the price comparison suggests.
The Invoice That the Context Window Missed
Baseten used a concrete failure scenario in the launch post to explain why the 128k context upgrade matters for decision quality. An AI agent approving invoices for payment checks the current invoice, confirms the amount matches the purchase order, and returns a yes. That logic is correct in isolation. But if the original purchase order was canceled three days earlier, that cancelation lives in an email thread or a CRM note, not in the invoice itself. At 16k context, the agent either does not see the email thread or receives a summary that may not preserve the specific cancelation signal. It approves a payment it should reject.
At 128k context, the invoice, the original order, the email thread flagging the cancelation, and the payment policy document all fit in a single call. The model decides with complete information rather than a compressed approximation of it. This failure mode is not unique to invoicing: it appears in content moderation (a post looks fine until you read the user's last 200 messages), in credit review (an application looks strong until you see the history of prior interactions), and in customer support routing (this complaint reads as routine until you see that the customer escalated twice before in the same week). The 8x context expansion addresses a structural limitation in how the prior generation handled any decision that required reading a document rather than a question.
The Risk Open Weights Cannot Solve: OpenAI Owns the Decision Moment
The bear case for bison-1's position is not about benchmark scores. It is about where decisions happen in production agent stacks. OpenAI's Decisions API is natively integrated with GPT-Live and, through the ChatGPT Wallet agent commerce workflow, with the checkout and payment layer for consumer-facing agents. When an agent browsing on a user's behalf reaches a decision point, that decision is embedded in the same session, not a separate API call to an external model. OpenAI controls the moment of action in its own ecosystem, and that ecosystem is where most consumer-facing agent applications run today.
Critics argue that the open-weight advantage depends on teams having the engineering capacity to self-host, integrate, and maintain a separate model in their agent stack, work that many production teams outsource to exactly the kind of hosted platform Baseten runs. Using bison-1 on Baseten's hosted API means trusting Baseten's uptime, latency guarantees, and pricing stability, which is a different kind of platform dependency from OpenAI's but still a dependency. The risk is that teams choose Baseten for cost and flexibility, but operational overhead erases the savings at volumes below the self-hosting break-even point, which varies widely by team size and infrastructure maturity.
The Gemma 4 backbone also introduces an upstream dependency. If Google changes Gemma's licensing terms or restricts access to new versions, Base Labs' ability to train future bison-class models on that foundation becomes uncertain. Baseten's post acknowledges the current release is a starting point for further research on American backbones, but that research trajectory is not locked in writing. For enterprise procurement teams with long planning horizons, that is a real consideration alongside the benchmark numbers.
Three Events to Watch Before the Next Index Update
The Artificial Analysis Decision Index v2 launched November 8 and had a point release to v2.1 within two weeks, suggesting the benchmark team is updating frequently as new models ship. Watch for a v2.2 that could include new entries from d1.2-ultra and Drex Max v1.1, which tied for the top position at 73 in the November 8 release before bison-1 entered. Both providers now have a visible score to beat. How quickly the competitive response arrives will signal whether the decision model market has quarterly update cycles or weekly ones.
Watch for an OpenAI Decisions API model update. The API went GA at $0.10/M with GPT-6 Luna on October 22. The changelog shows no model updates since, but DevDay 2026 introduced decision-focused tuning on the Luna model that was not finalized at launch. If OpenAI quietly swaps in an updated gpt-6-luna before the next index update, the benchmark comparison resets. OpenAI has done this before with the Responses API, updating underlying models on a named endpoint without changing the documentation headline.
The OpenAI Agent Builder retires on November 30, 2026, nine days from now. Teams that built non-code agent workflows in Agent Builder must migrate to ChatGPT Workspace Agents or the Agents SDK before that date. Many of those teams use Agent Builder as a no-code layer over Decisions API calls for routing and classification. As they evaluate replacements over the next nine days, they face a choice between staying within the OpenAI ecosystem or evaluating alternatives like Baseten's decision models for the components that do not require tight GPT-6 integration. That migration window is the first real test of whether bison-1's benchmark position translates into adoption outside Baseten's existing customer base.
Open weights matching proprietary API benchmarks at the same price is not a footnote to the decision model story. It is the story.
Key Takeaways
- bison-1 scores 76 on the AA Decision Index v2.1, topping 19 decision models with a Brier score of 0.14, one point better than the previous leader d1.2-ultra at 75.
- Both bison-1 and groundhog-1 carry 128k-token context windows, an 8x increase over Baseten's October models, enabling agents to pass full documents without summarization preprocessing that introduced accuracy errors.
- bison-1 is priced at $0.10 per million input tokens, matching OpenAI's Decisions API exactly, while groundhog-1 sits at $0.035 per million for cost-sensitive high-volume workloads.
- Open weights for both models are on Hugging Face under the Base Labs organization, allowing self-hosting, domain fine-tuning, and infrastructure-level auditability that proprietary API alternatives cannot offer.
- The OpenAI Agent Builder retires November 30, creating a nine-day migration window during which teams evaluating decision model alternatives will encounter bison-1's benchmark position for the first time.
Questions Worth Asking
- If bison-1 matches GPT-6 Luna on the AA Decision Index benchmark, does that parity hold on the specific decision types in your production stack, or is the benchmark's dataset distribution different from your use case?
- Soft-label training improves calibration on the index's training distribution, but does it transfer to genuinely novel decision domains like a healthcare intake form or a legal contract review where human annotators rarely disagree on training data?
- As OpenAI wires Decisions deeper into GPT-Live and ChatGPT Wallet, does benchmark parity stop mattering if the ecosystem lock-in is at the application layer rather than the model layer, and how would you know before you've already committed?