Squared: Spend Controls, Open Weights, and the Agent Plumbing Race
Most of this week was noise — Cannes ad demos, cloud gaming, IPO palace intrigue. But four things underneath actually matter for how you run AI in your organisation: OpenAI finally giving enterprises real cost visibility, a Chinese open-weights model closing the gap, agent infrastructure going GA, and security research that should worry anyone deploying research agents. Here's the signal.
OpenAI ships enterprise spend controls and usage analytics
OpenAI added spend controls and usage analytics to ChatGPT Enterprise — letting organisations cap and see where their AI budget actually goes.
Why it mattersIf you've rolled out ChatGPT Enterprise and can't answer "who's spending what, on what", that's been a governance gap. This closes it. The decision to revisit: do you actually have an owner accountable for AI consumption, or has it been quietly sprawling across teams? The tooling now exists — the org structure is on you.
GLM-5.2 is the strongest open-weights text model yet
Chinese lab Z.ai released GLM-5.2 under an MIT licence — described by Simon Willison as probably the most powerful text-only open-weights LLM available.
Why it mattersThe gap between frontier closed models and what you can run yourself keeps narrowing. For workloads with data-residency or cost constraints, self-hosting a near-frontier open model is now a credible option — not a compromise. Worth a serious look before you sign another multi-year frontier-API commitment for everything.
Agent infrastructure goes generally available
Amazon Bedrock AgentCore harness hit GA — two API calls to define and run a sandboxed agent — while open-source meta-harnesses like omnigent let you swap between Claude Code, Codex and Cursor without rewriting.
Why it mattersThe plumbing for production agents is maturing fast, and the meta-harness pattern is the smart move: don't hard-wire your stack to one provider. If you're building agents, design for harness portability now — switching costs are the trap you want to avoid.
Research agents can leak your secrets
ServiceNow's MosaicLeaks work tested whether research agents keep confidential data confidential — and a Beyond Static Leaderboards paper found no single benchmark covers more than a handful of the dimensions real deployment exposes.
Why it mattersAgents that browse, retrieve and call tools are an exfiltration risk most procurement processes don't test for. Before you let an agent touch sensitive data, ask your vendor what their leak testing looks like — and don't trust a leaderboard score as proof it's safe in your context.
Strip the IPO gossip and the marketing-conference theatre and the week's real message is consolidation: cost controls, open-weights parity, and agent infrastructure are all maturing at once. The two decisions worth your time — put a named owner on AI spend now that you can finally measure it, and don't lock your agent stack or your data to a single provider before you've tested it for leaks. Everything else this week was noise.
Working on something hard in AI? Just reply to the email — Daniel reads and answers every one.
Get Squared in your inbox each week
Free. The few things that actually matter, and what they mean for you.