GPT-5.6 in your office and Deutsche Telekom bets on AI

Today’s most important AI news

Saturday brought news of enterprise consolidation more than pure research. OpenAI keeps pushing corporate adoption on two simultaneous fronts: a new model installed in the workplace of millions of people, and a case study showing how a global-scale telco is rebuilding its operations with AI. There were no big controversies or disruptive launches, but what arrived today has direct implications for anyone working at a mid-size or large company.

GPT-5.6 now lives inside Microsoft 365 Copilot — and that changes your Monday

OpenAI confirmed that GPT-5.6 is now the preferred model inside Microsoft 365 Copilot, which means it’s active in Word, Excel, PowerPoint, corporate chat and Cowork. It’s not an announcement of future availability: the replacement has already happened. For the millions of users who access Copilot through enterprise licenses, the upgrade was silent and automatic.

GPT-5.6 isn’t GPT-5 with a cosmetic patch. OpenAI positions it as a model optimized specifically for productivity workflows — faster response times, better quality in synthesizing long documents and more precise reasoning on tasks that combine data with text. In practice, that translates into more coherent meeting summaries, Excel formulas generated with fewer errors and Word drafts that need less human editing.

For product managers and executives who have Copilot deployed across their teams, the relevant moment isn’t the announcement but the conversation that follows: are your teams using these capabilities or paying for a license nobody takes advantage of? The gap between having access to GPT-5-class AI and using it effectively remains the biggest adoption problem at Latin American companies.

Deutsche Telekom rebuilds a telco from the inside with OpenAI

Deutsche Telekom published, together with OpenAI, a detailed case study on how it’s transforming its operations using language models. The coverage goes beyond customer service — usually the generic use case that appears in every corporate presentation — and includes internal employee workflows, network operations and what they call “the future of voice”.

What makes this case relevant is the scale and depth. Telekom isn’t experimenting with a chatbot in a pilot division: it’s rebuilding processes in customer service, in the way its network engineers diagnose problems and in how employees access internal information. The network operations component is particularly interesting — using AI to anticipate failures or reroute traffic is a demanding technical domain where errors have real consequences.

For Latin American companies in telecommunications, financial services or any industry with critical infrastructure, this case sets a concrete precedent: AI is no longer just a user-experience layer, but a tool that touches the systems that keep the business running. The question their technology teams should be asking isn’t whether they’ll get to that point, but how much time they have before the competition does it first.

Attention profiling in PyTorch: what ML devs should know

Hugging Face published the third installment of its series on profiling in PyTorch, this time focused specifically on attention mechanisms — the central component of practically any modern language or vision model. The series is aimed at engineers who train or fine-tune models and want to understand where compute time really goes.

Attention profiling is a topic many developers avoid because it seems deep and impractical, but it has direct consequences on costs. Knowing where a model wastes GPU cycles during training or inference can mean the difference between a run that costs $200 and one that costs $800. For teams with tight budgets — which in Latin America is the norm, not the exception — that efficiency isn’t a technical luxury but an operational necessity.

If you have engineers working with their own models or fine-tuning open models, this Hugging Face series is concrete, free reference material. It’s worth reviewing before your next training run.

In a nutshell

Today’s pattern is clear: AI is ceasing to be a conversation about possibilities and becoming a conversation about implementation. GPT-5.6 is already on corporate desktops without anyone having had to make an active decision, and Deutsche Telekom is showing that real transformation doesn’t happen in an isolated department but across the whole operating chain. What this suggests for the coming months is that the debate will no longer be “should we use AI?” but “why aren’t we getting results?” — and that’s a much harder question to answer.

---

Sources used: https://openai.com/index/gpt-5-6-preferred-model-microsoft-365-copilot, https://openai.com/index/deutsche-telekom, https://huggingface.co/blog/torch-attention-profile