Rules, Prices, and Plumbing

Three things hardened this week. Europe's AI transparency rules became enforceable law with a deadline four days out, Anthropic cut the price of frontier intelligence in half and quietly published a number showing how often it still fumbles ordinary office work, and the open standard that connects AI to your business software got rebuilt for production traffic. This was the week AI stopped being a demo and started behaving like infrastructure — with rules, prices, and plumbing to match.

The Bottom Line

Europe just turned "we never said it was a bot" into a fineable offense. Anthropic halved the cost of frontier intelligence, then published a benchmark showing its best model finishes barely a quarter of real business automations without help. And the wiring that connects AI to your CRM stopped being a custom project and became a standard your identity provider already understands. The businesses that win the next six months won't shop for the smartest model — they'll aim a narrow one at a single workflow, disclose it, log it, and let it run.

See what we build →

Top 3 Stories This Week

01 · European Union · AI Act · Transparency

Four Days Until Your Chatbot Has to Confess ].

On Monday, July 27, the EU's Digital Omnibus on AI entered full legal force. Four days from today, on August 2, Article 50 of the AI Act becomes enforceable — and it asks for one plain thing. Any AI system that speaks to a person in the EU must say it's a machine from the start of the first interaction, unless that's already obvious. Not buried in the terms of service. At hello. The same rule covers AI-generated images, video, and published text on matters of public interest: label it. The reach is what catches people. The obligation follows the user, not the company — a twelve-person shop in Ohio with EU customers on its contact form is inside the fence. The European Commission last updated its guidance on July 20, leaving businesses under two weeks to read it and act. Penalties reach €15 million or 3% of worldwide annual turnover, whichever is larger, though the Commission says regulators must weigh proportionality for small and mid-size firms. One reprieve: systems already on the market get until December 2 for the machine-readable marking rule. The disclosure itself starts Sunday.

Microservice Opportunity: Build a disclosure-and-logging layer that sits in front of every bot you run — website chat, email autoresponder, AI-drafted posts. It injects the "you're talking to AI" notice and writes a timestamped record proving the notice fired. The notice takes an afternoon. The audit trail — the part that answers a regulator's question a year from now — takes a week. Retrofitting it after an inquiry costs far more than either.

02 · Anthropic · Claude Opus 5 · Automation

The World's Best Model Still Fails 3 of 4 Office Tasks ].

Anthropic released Claude Opus 5 on July 24 at $5 per million input tokens and $25 per million output — the same price as the model it replaces, and half what the top-end Fable 5 charges. It more than doubles its predecessor's score on Frontier-Bench and lands within 0.5% of Fable 5 on coding work at half the cost per task. Then comes the number nobody put in the headline. Zapier built AutomationBench out of real business automations — the unglamorous kind that pull a form submission into a CRM and fire the right follow-up email. Anthropic reports Opus 5 passing roughly 1.5× the rate of the next-best model at the same cost per task. Independent trackers put the raw number at 26%, against 17% for both its predecessor and Fable 5. Read that twice. The best score anyone has posted is also a 74% failure rate on the exact work owners keep being told AI will take off their plate. Frontier intelligence is now cheap. Turned loose and unsupervised, it is not yet dependable.

Microservice Opportunity: A 26% pass rate across open-ended automation is a science project. A 95% pass rate on one narrow task is a hire. The difference is scoping, not model choice. Pick the job you repeat forty times a week — quote follow-ups, invoice coding, review triage — and build one microservice that does only that, routing exceptions to a human. Opus 5's price makes the math work. The narrow scope makes the reliability work.

03 · MCP · Integrations · Enterprise Auth

The Wiring Between AI and Your Software Got Rebuilt ].

On July 28, the Model Context Protocol — the open standard that lets an AI assistant reach into your business software, from your bookkeeping to your ticket queue — shipped its largest revision since launch. Its developer kits now pull close to half a billion downloads a month, and the protocol went stateless: every request stands on its own, so a server scales behind an ordinary load balancer instead of nursing a live connection per user. The bigger change is quieter. MCP now speaks proper OAuth 2.0 and OIDC, the same login machinery behind Okta and Microsoft Entra. Your AI tool signs in as an identity your IT already governs, with the same permissions and the same audit trail as the employee who'd otherwise do the work. Microsoft, Google Cloud, AWS, Cloudflare, and Figma shipped support the same day, and a new twelve-month deprecation policy means what you build this quarter still runs next year. The expensive, brittle part of an AI project just got boring. Boring is what you want in plumbing.

Microservice Opportunity: "It isn't secure enough to let AI touch our CRM" just lost its footing. Wrap one system — scheduling, inventory, ticketing — in an MCP server that authenticates through the identity provider you already pay for. The AI reads and writes with exactly the permissions of the person it stands in for, and every action lands in a log you can hand an auditor. Build the first one carefully; the second costs a third as much.

This week in numbers: €15M — maximum Article 50 fine, or 3% of worldwide revenue, whichever hurts more, enforceable August 2 · 26% — Claude Opus 5's pass rate on Zapier's real business automations, best in the world and still a 74% miss · 500M — monthly downloads of MCP's developer kits, now that the standard runs stateless and speaks enterprise login

Sources

TL;DR — Key Takeaways

Top 3 Stories This Week

Four Days Until Your Chatbot Has to Confess ].

The World's Best Model Still Fails 3 of 4 Office Tasks ].

The Wiring Between AI and Your Software Got Rebuilt ].

Related answers