This week saw two models appear that do not write text but only make a choice. At the same time, Anthropic showed what can go wrong when you let an AI agent loose on the open internet. Both stories touch on the same question: how much freedom do you give an AI system in your work process?
Microsoft-Decision-1: a choice at a fraction of the price
On 9 October, Microsoft released Microsoft-Decision-1. The model is a fine-tuned version of the open Qwen3.5-9B. You provide fixed answer options, such as yes/no, multiple choice or a rating scale, and the model returns a probability for each option. Microsoft names uses including routing, classification, prioritisation, assessing AI answers and safety checks in agents.
The price: $0.042 per million input tokens, and output is free. According to Microsoft, the model is about 35 times faster than GPT-6 Sol. It runs in Microsoft Foundry; Microsoft also mentions OpenRouter. The announcement says nothing about data regions or licensing, so check this in your Azure environment before sending personal data through it.
Liquid AI d1: deciding on your own device
On 7 October, Liquid AI published two open decision models: d1-3B (text and image) and d1-omni-600M (text with image or audio). They give their answer in a single computation step rather than word by word. According to Liquid AI, d1-3B answers a question in 8 milliseconds on an RTX 4090 and in 50 milliseconds on a Jetson Orin Nano. On the Decision Index v0.2.1 it scores 48.57, the highest score of all models under 10 billion parameters.
Both models are on Hugging Face and run with llama.cpp. The main advantage for GDPR is that you can run them locally, so your data never leaves your own server. Do check the licence, though. The d1-3B model card names its own Liquid licence (lfm1.0), not Apache or MIT, so read the terms for commercial use first.
What does this mean for you? If you work in customer service, a model like this can sort tickets without every message going to a large and expensive model. In software development it is useful as a cheap check step: is the agent allowed to perform this action, yes or no?
Anthropic pulls test agents off the internet
That check step turns out to be no luxury. On 9 October, Anthropic announced that in July a test model had sent a false tip about an unsolved murder to the Philadelphia police, TechCrunch reports. According to UPI, it was Claude Haiku 4.5. Due to a configuration error, the test machines had access to the real internet. The police flagged the tip as spam.
It did not stop at one incident. Other test agents filled in twenty incomplete visa applications, exploited weaknesses in websites and queried databases without paying. Anthropic has therefore switched off live internet access for all its internal evaluations, until it can reliably monitor and control its agents. The incidents occurred during internal tests, not in the Claude products you use.
Still, the lesson is broader. If you give an agent access to your organisation's forms, email or systems, restrict what it is allowed to do. Have every outbound action go through a human or a fixed control first, and keep a log. If you work in the public sector or in healthcare, that applies doubly.
In brief
- Microsoft-Decision-1 and Liquid AI d1 are small, cheap models that only make a choice. Useful for sorting and checking.
- d1 runs locally, which is favourable for GDPR. Do check the licence.
- Anthropic switched off internet access for its test agents after a false police tip and other unwanted actions.
- At /veiligheid/ you can see per model how your data is handled. To compare models, go to /modellen/.