AI Daily Briefing for Wednesday, May 6th, 2026.

Anthropic made the biggest splash this week, launching ten ready-to-run Claude agent templates for financial services. These cover pitchbook creation, KYC screening, month-end closes, earnings reviews, and more — with deep integrations into FactSet, S&P Capital IQ, PitchBook, Dun & Bradstreet, and Moody's. It is not just about the agents. These templates ship with governed, real-time data connectors, MCP apps that embed partner tools directly inside Claude, and Claude add-ins for Excel, PowerPoint, Word, and Outlook. The Excel add-in can build financial models from filings, the PowerPoint add-in drafts decks that auto-update when numbers change, and Outlook acts as a chief of staff that triages your inbox. According to X analyst reports, the announcement sent Thomson Reuters stock down five percent, and FactSet also took a hit — investors are pricing in real disruption to legacy financial data incumbents. On the benchmark side, Anthropic says Claude Opus 4.7 leads Vals AI's Finance Agent benchmark at 64.3 percent.

Not to be outdone, OpenAI's Greg Brockman announced ChatGPT for Excel and Google Sheets on the same day. The timing was clearly a counter to Anthropic's Office suite push, and it signals that both frontier labs now see the enterprise productivity layer as prime real estate for AI adoption.

Google had two stories in play. First, a privacy firestorm: multiple reports and a detailed forensic analysis by That Privacy Guy confirm that Chrome has been silently downloading a four-gigabyte Gemini Nano model to users' devices — without consent, without an opt-out checkbox, and with automatic re-download if users delete the file. This story hit over thirteen hundred points on Hacker News. The analyst argues it breaches the EU ePrivacy Directive and GDPR, and at Chrome's two-billion-device scale, the one-time climate cost of that model push alone is estimated between six thousand and sixty thousand tonnes of CO2 equivalent.

On a more positive note, Google released Multi-Token Prediction drafters for the Gemma 4 family, delivering up to three times inference speedup with zero quality degradation. These speculative decoding drafters are open-sourced under Apache 2.0 and available today on Hugging Face, Kaggle, Ollama, and vLLM. They work by pairing a lightweight drafter model with the main Gemma 4 model — the drafter predicts several tokens at once, and the target model verifies them in parallel, dramatically reducing memory-bandwidth bottlenecks.

In the Apple ecosystem, reports indicate that iOS 27 will let users swap rival AI models across the operating system, with a fall 2026 release. This would allow third-party models like Claude or Gemini to replace Apple's own AI in Siri, writing tools, and other system-level features — all while maintaining Apple's privacy architecture. It's a big step toward model pluralism on mobile.

In legal and copyright news, a court filing unsealed yesterday alleges that Mark Zuckerberg personally authorized and encouraged Meta's copyright infringement in training its AI models. The publishers' lawsuit, reported by Variety, claims internal Meta documents show leadership knowingly used pirated datasets. This joins the growing pile of training-data litigation that could reshape how foundation models are built.

Harvey, the legal AI company, launched over five hundred use-case-specific agents plus an Agent Builder for law firms to create custom workflows. Early access is now open, and it represents one of the most ambitious vertical-agent plays in the legal sector.

On the infrastructure side, Cloudflare and Stripe launched a protocol that lets AI agents create Cloudflare accounts, start paid subscriptions, register domains, and deploy code — all programmatically, with humans in the loop only for terms-of-service acceptance. It effectively makes agents first-class cloud customers, and comes with one hundred thousand dollars in Cloudflare credits for Stripe Atlas startups.

Finally, Coinbase CEO Brian Armstrong announced a fourteen percent workforce reduction, citing the need to flatten the organization and accelerate AI-native operations. And a widely-shared analysis from Reflex found that computer-use agents are forty-five times more expensive than structured API calls for the same tasks — a reminder that while agentic AI is compelling, the economics still heavily favor structured integration.

That's the briefing for May 6th. I'm Bob — back tomorrow.