Squared: The Copilot Model Swap and the Agent Land-Grab

OpenAI released the GPT-5.6 family — three sizes (Luna, Terra, Sol) priced from $1/$6 to $5/$30 per 1M input/output tokens per Simon Willison — and it's now the 'preferred model' powering Copilot across Word, Excel, PowerPoint and Chat, even as breakup chatter between the two firms continues (TechCrunch).

If you run Microsoft 365, your Copilot behaviour just changed without you choosing it — outputs, latency and cost profile will shift. Re-run your key Copilot workflows this week and check nothing regressed. And note the 'preferred', not 'exclusive' language: the Microsoft-OpenAI relationship is loosening, so don't architect anything on the assumption it's permanent.

OpenAI launched ChatGPT Work (a rebrand of Codex) — an agent that takes action across apps and files and can run 'for hours if needed'. Desktop Work can touch local files and apps with permission; cloud and desktop conversations don't yet sync (Ars Technica, Simon Willison).

An agent with hours of autonomy and access to local files is a governance question, not a productivity toy. Before anyone in your org points this at real data, decide what it's allowed to touch and how actions are logged. The permission model is the whole risk surface — treat it like handing a contractor your file server.

OpenAI is sunsetting ChatGPT Atlas, the agentic browser it announced in October, folding the ambition into ChatGPT Work (The Verge).

A reminder to stop betting your workflows on any single AI vendor's shiny standalone product. Things launched with fanfare get retired in months. Build on the durable primitives — the models and APIs — not the app-of-the-quarter, and keep an exit path for anything you adopt.

Ars Technica reports the NYT alleges OpenAI faked an inability to search training data and hid or deleted billions of ChatGPT logs — potentially exposing the company to court sanctions.

This is the legal cloud over the model you may be standardising on. It won't stop you using GPT-5.6, but it should inform your contractual posture: data-retention terms, indemnities for copyright claims, and a fallback provider. Concentration risk on one vendor cuts both ways.

Microsoft's 2026 sustainability report shows emissions rose 25% in 2025 to 34 million metric tons, driven largely by datacentre expansion (The Verge, via GeekWire).

If you have public net-zero commitments, heavy AI usage now has a measurable carbon line item that your ESG reporting will have to own. The 'AI is free once it's in the cloud' assumption is wrong — someone is paying in power, and increasingly that scrutiny lands on customers too.

Trending this week: open connectors linking 1000+ SaaS apps to agents (open-connector), @agent mentions routing to Codex and Claude Code in Slack (opentag), and an autonomous red-teaming harness (T3MP3ST) at 4,200+ stars.

The ecosystem is racing to give agents access to everything you use — and to attack it. If your teams are experimenting with these, the offensive-security tooling is trending for a reason. Assume agents will be probed, and make sure someone owns adversarial testing before agents touch production.