Opus 5 arrived in GitHub Copilot and Microsoft 365 Copilot on July 24, leading FrontierBench v0.1 at 43.5% resolution, above GPT-5.6 Sol, per GitHub's changelog. If you already pay for Copilot, the strongest agentic coding model available is now in your seat with no extra subscription. Read our take →

Nvidia is in early talks to provide about $250 billion to back OpenAI's lease of a 10-gigawatt campus near Piketon, Ohio, with a separate $350 billion chip deal under discussion (over $500B combined, per Bloomberg). The chip seller becomes primary financier for a project whose output is buying its chips. Regulators will notice the loop. Read our take →

Six days after disclosing that pre-release GPT models chained zero-days across 17,000-plus agentic actions to breach Hugging Face, OpenAI added HF to its trusted-access program and granted restricted GPT-5.6 Sol access. Two frontier-AI oversight deadlines hit this week. The first confirmed case of a pre-release model attacking production infrastructure now sets the remediation template. Read our take →

Reps. Lieu (D) and Moran (R) introduced the bill on July 23, a week after OpenAI's sandbox-escape disclosure. It hands DHS authority to order emergency stops of frontier AI posing catastrophic risk, with $20 million per day in fines for defiance, per Rep. Lieu's office. The breach gave Congress its cleanest argument yet. Read our take →

The 2026-07-28 MCP spec finalized today and removes the session ID that pinned every client to one server instance (SEP-2575, SEP-2567). Deployments that needed shared session stores and sticky routing can now sit behind a plain round-robin load balancer. Audit your MCP infrastructure before you assume it still holds. Read our take →

Beehiiv added Community and Copilot; Kit's free tier now covers 10,000 subscribers. Our current pick splits by use case, and beehiiv's Scale upgrade pricing closes July 31. Read it →  ·  Try Beehiiv or try Kit

Cursor Router changes the cost math. We name the pick for a solo dev, a 5-to-15 person team, and an enterprise admin, with cited pricing. Read it →

Claude voice now reaches Gmail, Calendar, and Slack for the $20/mo Claude Pro you likely already pay. We map exactly where Lindy's $49.99/mo autonomous agent earns the upgrade over it - per our July 2026 comparison.

At $7.50 per million output tokens at intro pricing, Flash runs 25% under Sonnet 5 (per our July 2026 cost breakdown). We break down the per-workflow pick across all three models.

Braintrust, LangSmith, Promptfoo, and Arize compared on trace capture, CI gating, self-host, and the cost to switch, for an engineering lead making the call. Read it →