AI briefing, week 41/2026: what mattered for SMEs

What happened. On 1 October 2026 Microsoft announced that new Microsoft 365 Copilot Business licences bought through a partner (CSP) will come with additional usage-based billing switched on by default from 1 December 2026. The fixed licence stays; usage for certain services is billed on top. The original date was 2 November. The default spending limit is 4,000 Copilot Credits per user per month, and admins can change it. This billing covers extra services such as Copilot Cowork and the Work IQ APIs. The change will not reach Germany, France, Italy, Spain and several other countries at first. Austria is not on that list. (Microsoft Partner Center, October 2026 announcements)

What it means for an SME. With 20 Copilot users, the default limit allows up to 80,000 credits a month before anything stops. What that costs in euros depends on the price per credit in your agreement. One default also matters: spending policies apply automatically to future services Microsoft adds, unless someone turns off "Auto-apply new services". (Microsoft Learn, usage-based billing)

Our recommendation: act (Austria). Ask your Microsoft partner whether your next Copilot order falls under the new rule. Set a limit that fits your budget, and name the person who reviews usage under "Cost management" in the admin center. In Germany: watch.

What happened. On 30 September 2026 Google announced skills: reusable instructions for Gemini that work in the Gemini app and in most Workspace apps. The Workspace rollout starts on 5 October 2026 for Rapid Release domains, and the remaining domains follow by mid-November. On 17 November 2026 Gems move into the Gemini app settings. No sooner than 1 March 2027, business and enterprise customers can no longer create, edit or use Gems. Remaining Gems in the Gemini app automatically become inactive skill drafts there. The person who created them must review and activate each skill before using or sharing it. Skills in the Gemini app and skills in Workspace do not sync: for Workspace you must recreate them in Google Workspace Studio. Admins control in the Admin console who may use skills in Workspace. (Google Workspace Updates, Google Workspace: transition from Gems to skills)

What it means for an SME. If your team uses Gems for fixed routines, such as draft quotes or a house writing style, those routines will depend on skills that someone must activate or recreate in Workspace. For Gems owned by former employees, decide in advance how to keep the instructions and recreate them as skills.

Our recommendation: plan. List the Gems your team actually uses and give each one an owner who activates it as a skill or recreates it in Workspace Studio. Check in the Admin console that skills are allowed for the intended users, and test the most important routines before 1 March 2027. General rules for AI tools in a team are in our AI usage policy template.

What happened. On 5 October 2026 OpenAI announced that supported text from ChatGPT and Codex in the EU will carry an invisible watermark over the coming weeks, for eligible users on all plans. In the API, the watermark is available now for selected models worldwide, as an opt-in that is off by default. Access to the detection tool goes first to approved researchers and expert organisations only. The background is the transparency duties of the EU AI Act. (OpenAI, coverage by TechCrunch)

On 29 September 2026 OpenAI also released GPT-6.1 Sol (OpenAI API changelog). For standard processing with up to 272,000 input tokens, the API price is the same as for GPT-6 Sol: 2 US dollars per million input tokens and 10 US dollars per million output tokens. Cached input costs 0.10 US dollars, compared with 0.20 for GPT-6 Sol. The larger GPT-6 Astra costs 10 and 50 US dollars. Longer prompts and other processing tiers have different prices. For endpoints that process data in a specific region, such as the EU, OpenAI charges a 10 percent uplift on models released on or after 5 March 2026. (OpenAI API pricing)

What it means for an SME. The watermark is a technical signal. It does not prove that a text is correct or that a labelling duty is met. According to OpenAI, detection can fail after a text is edited. If your automation uses an EU regional-processing endpoint for an affected model, include the 10 percent uplift.

Our recommendation: plan. State in your internal policy that AI text is checked before use, whatever the watermark says. If you run automations on GPT-6 Astra today, you can test GPT-6.1 Sol with 20 of your own cases before you switch.

What happened. On 6 October 2026 Mistral released Mistral Large 4 as a public preview in its API: a multimodal model with about one trillion parameters and a context window of one million tokens. During the preview it runs in Mistral's own data centres in Europe, with 50 percent off for the first two weeks. Open weights are due by the end of October. (Mistral, Mistral changelog)

On 3 October 2026 Aleph Alpha from Heidelberg (Aleph Alpha) released Kolibri: a model for German and English with 78 billion parameters, about 3.5 billion of them active, under the Apache 2.0 licence. According to the model card, the FP8 model weights take about 78 GB; the listed minimum hardware includes two Nvidia A100 80 GB cards or one H200. Aleph Alpha describes it for workflows in which a person reviews the output before anyone acts on it. (Hugging Face, Aleph-Alpha/Kolibri-1)

In the same period Mistral retired OCR 4.0 on 30 September 2026. The successor is OCR 4.1 at the same price. (Mistral changelog)

What it means for an SME. Companies that cannot or do not want to send data to US providers have more choice. With its open licence and focus on German, Kolibri is a candidate for running AI locally, but it needs powerful GPU hardware, which you can run yourself or rent from a provider in an EU data centre. Mistral Large 4 is currently a public preview; check the applicable terms before production use. The model card lists 1.36 US dollars per million input tokens and 4.18 US dollars per million output tokens, and the launch discount lasts two weeks. (Mistral model card)

Our recommendation: watch, and act on Mistral OCR. If you use Mistral OCR in invoice or document processing, check now whether the workflow still calls the model mistral-ocr-4-0. For the new models: test with your own cases and check the data terms before you process customer data. When local operation pays off is covered in Local AI or EU cloud.

We also checked Meta, DeepSeek and Qwen and included no item from them in this issue. Left out on purpose: reports about IPO plans and valuations of AI companies, and leaderboard comparisons. They do not change your workflows. Only a test with your own cases shows which model fits your tasks; our comparison ChatGPT, Claude or Gemini covers the business plans.

If you want to know what one of these changes means for your existing workflows, we are happy to talk it through in an AI consulting call.