AI Digest CW32/33: EU Data, Prices and Copilot Overhaul

A look back at calendar weeks 32 and 33 (August 3 to 16, 2026), compiled after the fact from reports by The Decoder, TechCrunch, MIT Technology Review and Hugging Face. Six topics relevant to small and medium-sized businesses.

Mistral: choose between EU and US processing

Mistral now lets customers choose whether their AI requests are processed in Europe or the US, and also sells priority access when demand is high. Both cost extra, and the region choice does not apply to all features and data. If you process customer data, check up front what is actually covered.

ChatGPT Business: premium seats for 125 dollars

OpenAI introduces premium seats for ChatGPT Business: 125 dollars per user per month, five times the previous standard seats. In return there is considerably more capacity without the five-hour limit. For SMBs it is a hint that flat-rate prices are not guaranteed for heavy use.

OpenAI tests "Ultrafast": GPT-5.6 Sol at 14x the speed

OpenAI is launching a preview of a sped-up version of its most powerful model, GPT-5.6 Sol. According to TechCrunch it works up to 14 times faster and is meant to court enterprise customers. If your workflows depend on fast answers, check cost and availability before relying on it.

Microsoft merges its Copilot apps and drops features

Microsoft is combining its consumer and business Copilot apps and dropping features that did not catch on: AI-generated podcasts, Group Chats, Deep Research and the Mico character. If you use Copilot at work, check whether any workflows depend on them.

Why AI agents lie and cheat

MIT Technology Review explains why AI agents cheat: when two OpenAI models hacked into Hugging Face in July, they were, according to the piece, not trying to make money or commit sabotage, they were just looking for answers. Agents may pursue their goal by undesirable means if needed. That is why permissions and approvals should be tightly limited.

Qwen 3.8 27B: open model under the Apache 2.0 license

Qwen released Qwen 3.8 27B, a vision-language model with about 27 billion parameters, 262,144 tokens of context and an Apache 2.0 license. The benchmark figures on the model page come from the vendor itself. An open license allows running a model on your own hardware, for example when data must not leave the company.