Claude under the microscope: Anthropic opens the black box of LLMs

Today’s most important AI news

Today was a day marked by interpretability: Anthropic published a finding that changes the conversation about what really happens inside a language model when it processes a question. At the same time, the race to integrate AI into productivity tools advanced with GPT-5.6 as the base model in Microsoft 365 Copilot, and Deutsche Telekom presented the most detailed case yet of a telco reinventing itself around OpenAI. It’s a Friday dense with signals about where the industry is heading.

Anthropic opens the black box: what it found inside Claude

For years, the main criticism of language models wasn’t their performance but their opacity: nobody knew exactly what they processed or how they arrived at an answer. Anthropic has just changed that with a tool called the Jacobian lens, which let them observe Claude’s internal states in unprecedented detail while it solves tasks. What they found ranges from the expected to the unsettling.

The central finding is that Claude doesn’t simply predict tokens in sequence: there’s an intermediate space — not visible in the output — where the model seems to “deliberate” over concepts before generating text. Anthropic describes this as a hidden reasoning space that doesn’t correspond one-to-one with natural language. That has direct implications for those using Claude in production: the model may be processing alternative paths that never appear in the final answer.

For developers and product teams building on Anthropic models, this is relevant on two levels. First, it suggests that prompting techniques that try to “see” the model’s reasoning — like chain-of-thought — capture only a fraction of what happens internally. Second, it opens the possibility that Anthropic will develop diagnostic tools that make it possible to detect when a model is taking a problematic reasoning path before the error reaches the user.

GPT-5.6 takes over Microsoft 365 Copilot

OpenAI quietly updated the engine that powers Microsoft 365 Copilot: GPT-5.6 is now the default model in Word, Excel, PowerPoint, Chat and the Cowork feature. It’s not a launch with a press conference, but it carries real operational weight for any company already paying for Copilot.

GPT-5.6 isn’t GPT-5 with a minor patch. OpenAI presents it as a version optimized specifically for enterprise workflows: better reasoning over long documents, greater consistency on repetitive tasks and faster responses within the Office suite. For a product manager who uses Copilot to summarize meetings or an executive who generates reports with the model’s help, the change should be noticeable without needing to modify any configuration.

What’s worth monitoring is the speed at which OpenAI is iterating on versions. In less than a year we went from GPT-4 Turbo to GPT-4o to GPT-5 and now GPT-5.6 as the standard in enterprise production. For organizations with critical flows tied to specific model behaviors, this update cadence forces continuous evaluation — approving the model once and forgetting about it isn’t enough.

Deutsche Telekom rebuilds a telco around AI

OpenAI published the most extensive case to date of a telecommunications company reorganizing its operations using its models. Deutsche Telekom — with more than 200 million customers in Europe and the United States — describes a transformation that touches four layers: customer service, internal workflows, network operations and the future of the voice channel.

The most significant detail of the case isn’t the volume of automation but the focus on network operations. A telco using AI to answer tickets or summarize internal documents is no longer news. One applying language models to reason about network failures and anticipate degradation in critical infrastructure is a different leap, with reliability implications that go far beyond operational efficiency.

For entrepreneurs and executives in Latin America, the Deutsche Telekom case works as a roadmap. The region’s telcos face similar pressures — high operating costs, saturated customer service, complex networks — but with more limited resources. The pattern OpenAI describes, starting with the contact center and scaling toward network operations, is reproducible on a smaller scale. The question Latin American telcos should be asking is whether they’d rather learn from this case now or replicate it under competitive pressure in two years.

In a nutshell

The day’s pattern is the quiet maturation of AI in production. Anthropic’s finding about Claude’s internal states isn’t an academic curiosity: it’s the technical foundation for building more auditable models, something regulators and security teams will increasingly demand. At the same time, GPT-5.6 in Copilot and the Deutsche Telekom case show that enterprise integration is no longer in the pilot phase — it’s in the infrastructure phase. The frontier moving today isn’t benchmarks but operational trust: who can explain what their model does, and who can prove it works in real production for months. That’s where the industry’s next stage will be won or lost.

---

Sources used: MIT Technology Review — https://www.technologyreview.com/2026/07/09/1140293/anthropic-found-a-hidden-space-where-claude-puzzles-over-concepts/, OpenAI Blog — https://openai.com/index/gpt-5-6-preferred-model-microsoft-365-copilot, OpenAI Blog — https://openai.com/index/deutsche-telekom