---
title: The new models are not for chatting. They are for deciding.
description: Thinking Machines, Baseten, Nous and a research team all shipped models built to make choices and take actions, and the early numbers say the small, specialised ones hold their own.
canonical: https://hangingcontext.com/stories/2026-11-21/decision-models-week/
last-updated: 2026-11-21
url: https://hangingcontext.com/stories/2026-11-21/decision-models-week/
published: 2026-11-21T00:44:08Z
edition: 2026-11-21T00
newsroom: artificial-intelligence
beat: models-labs
trust: confirmed
source_count: 4
cited_manifest_ids: ['1795213500226621606', '1795215927298634098', '1795200130070535516', '1795199100960526099']
---

# The new models are not for chatting. They are for deciding.

*Thinking Machines, Baseten, Nous and a research team all shipped models built to make choices and take actions, and the early numbers say the small, specialised ones hold their own.*

Confirmed · 4 Manifests · Artificial Intelligence · Published 2026-11-21 00:44 UTC

## Key facts

- Thinking Machines released Inkling 2 Mini, a 26-billion-parameter open-weights model with 5.4 billion active parameters that matches Qwen3.8 27B on the Artificial Analysis Intelligence Index at about six times lower cost per task. [cites: 1795213500226621606]
- Baseten says its bison-1 decision model ranks first among 19 decision models on the Artificial Decision Index with a score of 76, priced at $0.10 per million tokens, with groundhog-1 at $0.035. (self-reported) [cites: 1795215927298634098]
- In a preprint, System 1 agents reached 97.0% accuracy on 1,952 requests to 157 biomedical tools, beating the best of 18 local LLM-agent setups at 95.2%, with no invented content in any call. (self-reported) [cites: 1795200130070535516]
- Nous Research launched Hermes Enterprise on Microsoft Foundry, giving agents their own Microsoft Entra accounts governed through the Microsoft 365 admin center, in preview for Frontier program companies. [cites: 1795199100960526099]

## What shipped

Thinking Machines released Inkling 2 Mini, an open-weights model that loops twelve of its twenty-four layers three times per token. The model totals 26 billion parameters but only 5.4 billion are active at a time, and Artificial Analysis reports it matches Qwen3.8 27B on its Intelligence Index while costing about six times less per task. It runs on a MacBook or a DGX Spark, carries a one-million-token context window, and ships under an Apache 2.0 licence. [cites: 1795213500226621606]

Baseten released two decision models, bison-1 and groundhog-1, built on Google's Gemma 4 backbone. The company says bison-1 ranks first among nineteen decision models on the Artificial Decision Index with a score of 76, and both models take 128,000-token contexts and multimodal input. Baseten prices bison-1 at ten cents per million tokens and groundhog-1 at three and a half cents, and has put the weights on Hugging Face. [cites: 1795215927298634098]

The common thread is what these models are for. Inkling 2 Mini is built around repeated passes over its own layers, and Baseten's pair are described as decision models rather than chat models. Neither release leads with conversation quality. Both lead with cost, size and a job to do. [cites: 1795213500226621606, 1795215927298634098]

> **From Synorb** — Run the cited Manifests yourself. Every claim in this paper points to a Manifest id. Get a Synorb Key and read the Manifests behind it. Existing 72 Hour Agent Demo keys keep their issued limits. New keys require email OTP verification. An existing demo key lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons. [Get a key](https://synorb.com/connect/start)


## Decisions, not text

A preprint from the same week takes the idea further. System 1 agents replace generative calls with a one-pass decision model that turns tool requests into multiple-choice questions. In a test of 1,952 requests to 157 biomedical tools, the framework reached 97.0 percent accuracy, beating the best of eighteen local LLM-agent setups at 95.2 percent. None of its calls contained invented content, against 5.5 to 19.7 percent of the LLM calls. [cites: 1795200130070535516]

The point is not that generative models are finished. It is that a growing share of agent work is a choice among known options, and a model that only ever picks from a list cannot invent an argument that was never offered. That trade, less freedom for more reliability, is the same one Baseten makes with decision models whose calibration it improved by scaling the use of soft labels in training. [cites: 1795200130070535516, 1795215927298634098]

## Agents get their own accounts

Nous Research launched Hermes Enterprise on Microsoft Foundry, giving agents their own Microsoft Entra accounts instead of borrowing a human's login. IT teams can govern the agents through the Microsoft 365 admin center, and each agent session runs in an isolated virtual machine without storing long-term passwords or tokens. The deployment is in preview for companies in Microsoft's Frontier program. [cites: 1795199100960526099]

Nous says Hermes Agent is the top agent for productivity and coding on OpenRouter, and the Enterprise version adds a shared library of learned skills plus a tool for tracking agent costs and returns. The pitch is that an agent with its own identity is easier to audit, limit and switch off than one wearing a person's credentials. [cites: 1795199100960526099]

## What to watch

Taken together, the releases point at a split in the model market. One side keeps chasing general conversation. The other builds smaller, cheaper models for a defined job, a decision, a tool call, a shift's worth of work inside a company's own systems. Inkling 2 Mini's price per task and bison-1's ranking both argue that the second side can now compete on the numbers, not just on principle. [cites: 1795213500226621606, 1795215927298634098]

The open questions are practical. Baseten's ranking comes from its own announcement, and the Artificial Decision Index is a young benchmark. Hermes Enterprise is still in preview, and System 1's results come from a preprint on biomedical tools, not from production deployments. What matters next is whether these models hold up when the job is messier than the test. [cites: 1795215927298634098, 1795200130070535516, 1795199100960526099]

## What is still unresolved

- Whether bison-1's ranking on the Artificial Decision Index holds up under independent testing
- Whether System 1 agents' accuracy on biomedical tools transfers to other tool domains
- When Hermes Enterprise leaves preview and what it costs outside the Frontier program

## Decision card

- **Useful for:** Founders pricing agent products by the task; Teams choosing a model for tool use or decisions rather than chat
- **Not for:** Teams that need open-ended conversation quality above all; Buyers who need production references before adopting a new model
- **Compare against:** General-purpose chat models on cost per completed task; In-house tooling built on hosted frontier APIs

## Provenance ledger

| Manifest id | Title | Source | Claims used | Published |
|---|---|---|---|---|
| `1795213500226621606` | Thinking Machines has released Inkling 2 Mini, an open weights looped transformer | [artificial-analysis-research](https://artificialanalysis.ai/articles/inkling-2-mini) | 1795213706081624413, 1795213706082624414, 1795213706086624420, 1795213706088624423, 1795213706078624411, 1795213706080624412, 1795213706084624417 | 11-20 00:00 |
| `1795215927298634098` | Meet bison-1 and groundhog-1, frontier American decision models | [baseten-blog](https://baseten.co/blog/meet-bison-1-and-groundhog-1) | 1795216137314634763, 1795216137315634764, 1795216137318634767, 1795216137318634768, 1795216137313634762, 1795216137316634765, 1795216137317634766 | 11-20 00:00 |
| `1795200130070535516` | System 1 agents: a method for local, reliable and fast agentic AI | [huggingface-daily-papers](https://huggingface.co/papers/2611.21053) | 1795200368914536650, 1795200368917536658, 1795200368915536653, 1795200368913536647, 1795200368910536642 | 11-20 00:00 |
| `1795199100960526099` | Introducing Hermes Enterprise on Microsoft Foundry: AI coworkers that are easy to use and easy to govern | [nous-research-blog](https://nousresearch.com/hermes-enterprise-on-microsoft-foundry) | 1795199399179530513, 1795199399180530514, 1795199399182530517, 1795199399176530510, 1795199399183530518, 1795199399184530519, 1795199399178530511, 1795199399181530515, 1795199399182530516 | 11-20 00:00 |

## Run these Manifests yourself

Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.

1. **Get a Synorb Key.** Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool `synorb-get-demo-key` (no header needed), or REST `POST https://api.synorb.com/api/connect/demo-key` with a JSON body such as `{"client":"your-agent-name"}`. The key is the `api_key` field of the answer. Save the key. Existing 72 Hour Agent Demo keys keep their issued limits. New keys require email OTP verification. An existing demo key lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons.
2. **Read each cited Manifest by id.** REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer <your Synorb Key>`. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief.
3. **Check the claim.** Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them.

Cited Manifest ids:

- `1795213500226621606`
- `1795215927298634098`
- `1795200130070535516`
- `1795199100960526099`

```sh
for id in 1795213500226621606 1795215927298634098 1795200130070535516 1795199100960526099; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done
```

> **From Synorb** — Connect your agent over MCP. Add https://mcp.synorb.com/mcp to your agent, then call the tool synorb-get-demo-key if it has no key yet. It needs no email and no form. Keys are free for 72 hours, and new keys are limited per network. [How to connect](https://synorb.com/connect/start)

## Related stories

- [Prediction moves from benchmarks to real outcomes](https://hangingcontext.com/stories/2026-11-21/prediction-moves-to-real-outcomes/)
- [US startups run Chinese open-weight models in production — and pay a safety price](https://hangingcontext.com/stories/2026-11-20/us-startups-run-chinese-open-weight-models/)
- [Agents got a wallet this week. The spending limits arrived with it.](https://hangingcontext.com/stories/2026-11-20/agents-get-a-wallet/)
- [Superintelligence has a physical-world problem, and the labs know it](https://hangingcontext.com/stories/2026-11-19/physical-world-is-the-bottleneck/)

## Links

- HTML: https://hangingcontext.com/stories/2026-11-21/decision-models-week/
- JSON: https://hangingcontext.com/stories/2026-11-21/decision-models-week/index.json
- Markdown: https://hangingcontext.com/stories/2026-11-21/decision-models-week/index.md

Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify.

---
Reuse (terms v1): Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product. Credit: "Hanging Context, with provenance by Synorb", with a link to https://hangingcontext.com/stories/2026-11-21/decision-models-week/. Terms: https://hangingcontext.com/terms/reuse/
