This week in AI: 8 things you may have missed

Three weeks since the last recap and the agent layer left the browser. Anthropic skipped a flagship model at its big keynote and bet the quarter on orchestration. OpenAI made the default ChatGPT model meaningfully less wrong, plugged it into your bank account, and put Codex in your pocket. Here's the 90-second catchup so you can get back to building.

THIS WEEK IN AI

8 things from Anthropic + OpenAI

ANTHROPIC

• Code with Claude 2026 — no new model, all orchestration. At its May 6 SF keynote, Anthropic deliberately skipped a flagship model release and shipped Managed Agents GA, Remote Agents (start a session on your laptop, continue it from your phone), an Advisor tool for long-running plans, Claude Code routines, and CI auto-fix. CPO Ami Vora framed the bet plainly: the next quarter of advantage comes from chaining models, not benchmarking them.
• Claude for Small Business. A package of native connectors for QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365 — Claude now sits inside the tools 30M U.S. small businesses already pay for. SMB is suddenly a first-class distribution channel, not a "we'll get to it" segment.
• Big-4 alliances stack up. PwC expanded its strategic Claude alliance May 14. KPMG followed May 19 with a global deal across 276,000 employees and 138 countries. Enterprise lock-in is compounding fast on Anthropic's side of the chart.
• Boring infra wins: cache diagnostics + Claude Fast. Cache diagnostics shipped in public beta — pass diagnostics.previous_message_id and the API tells you exactly where your prefix diverged (cache_miss_reason). Fast Mode landed in research preview for Opus 4.7. Two unflashy upgrades that quietly take 30–50% off your bill if you tune for them.

OPENAI

• GPT-5.5 Instant becomes the default — and 52.5% less wrong. Rolled out May 5 to all ChatGPT users and the API. OpenAI's headline: 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in medicine, law, and finance. The "AI is unreliable" objection just lost a leg in sales calls.
• ChatGPT Personal Finance (Pro preview). Connect Chase, Schwab, Fidelity, Robinhood, Amex, Capital One — Plaid-powered, 12,000+ institutions, U.S. Pro tier only at launch (May 15). Ask it about your spending, plan for next year, get real account context in the answer. The "agent at the bank" wedge is officially open.
• Codex Mobile + Goal Mode GA + Appshots + remote computer use. Codex now ships on iOS/Android for Free and Go users. Goal Mode exited beta — give Codex an objective and let it work for hours or days. Press both Command keys on macOS to ship the active window to Codex with screenshot + text ("Appshots"). And Codex can drive your Mac after it locks, including from your phone, with short-lived authorization and a relock-on-input safeguard.
• Codex-Spark with Cerebras. First OpenAI × Cerebras model in the wild, research preview for ChatGPT Pro in the Codex app, CLI, and IDE extension. Text-only, 128k context. Anthropic shipped Fast Mode the same week. The speed-tier race is officially on.

---

FROM THE SITE

Three things to read after this recap:

• Claude Code on Your Phone — Written before Remote Agents existed, accidentally became the playbook for it. How to actually use Claude from your phone without losing your mind.
• Claude Code for Non-Technical Founders — The on-ramp for the audience Anthropic just opened with Claude for Small Business. If you're a founder who isn't a developer, this is your week.
• Startup Financial Dashboard — Weekend build the new ChatGPT Personal Finance product just validated. Plaid + AI + a clean weekly digest — solo founders are the wedge before consumers.

---

3 TAKEAWAYS FROM THE WEEK

• The agent layer left the browser. Remote Agents on Anthropic, Codex Mobile + remote computer use on OpenAI. The AI doesn't live in a tab anymore — it lives on your phone and reaches back into your laptop. If you're building dev or productivity tools, "where does the agent run?" is now a product question, not an infra one.
• Distribution beats capability — louder this week than ever. Claude inside QuickBooks. Claude across 276K KPMG employees. ChatGPT plugged into 12,000 banks. None of those are model improvements. They're acquisition channels. The "best model" stops mattering once the AI is already inside the workflow.
• Hallucinations got measurably better — bank the credibility. A 52.5% drop on high-stakes prompts isn't a benchmark flex; it's a sales-objection killer. If you sell anything AI-powered in regulated-ish spaces (legal, medical, finance, ops), update your one-pager this weekend. The skepticism wall just got 50% shorter.

---

Sneak peek at Monday…

A weekend MVP that turns your bank's CSV export and your last 30 transactions into a one-page "what should I cut?" brief — built in 4 hours with Claude + the new Plaid pattern.

---

PS — Which of these 8 was new to you? Hit reply with the number — I'm picking next week's deep dive based on what people missed. Last edition the top vote was Codex computer use; let's see what wins this round.