Meta Superintelligence chief Alexandr Wang told staff on July 2 that the in-training Watermelon model matches OpenAI's GPT-5.5 on key benchmarks, using an order of magnitude more compute than predecessor Avocado, per Benzinga. No release date, and no third-party numbers yet - read it as a compute flex, not a shipped win. Read our take.

The Apache 2.0 Lean 4 agent solved 587 of 672 PutnamBench problems at roughly $4 per problem, versus an estimated $300-plus per problem for Seed-Prover's high setting, per Mistral's own report via MarkTechPost. Open weights on Hugging Face. For formal-verification work, the cost curve is the news.

ByteDance's Doubao and Alibaba's Qwen are pulling features that let agents impersonate real people or adopt humanlike personas, ahead of China's synthesis-tech rules. If you ship an agent that takes a first-person user identity in China, you have nine days to get compliant.

OpenAI, Google, and Anthropic are finalizing voluntary release standards with the administration. EO 14409 already sets a classified benchmarking process and a 30-day government access window before public launch. Watch whether "voluntary" stays voluntary.

The tool we lean toward over Lindy when you want to see every step of a workflow fail. Visual canvas, step-by-step debugging, and a Zapier-compatible learning curve.

Returns clean markdown ready for agent and RAG pipelines - the managed alternative when WebMCP and Playwright are more plumbing than you want to maintain.

Two CVSS 9.8 zero-click flaws patched in 3.0, a split Teams price at $19 solo and $38 shared, and a new iOS app - our full July verdict on who should upgrade now.

Make runs a visual, debuggable workflow; Lindy acts like a person in your inbox. Decision table plus current pricing for solo operators picking one.

Sonnet 5 leads in-repo editing at 63.2% SWE-bench Pro vs GPT-5.5's 58.6%, priced at $2/M input through August 31. GPT-5.5 owns terminal agents; Gemini 3.1 Pro owns front-end. The per-persona pick.

Five controls Microsoft, OWASP, and TrueFoundry agree on, with per-tool notes for Cursor, n8n, and Firecrawl plus a copyable audit table.

WebMCP entered a Chrome 149 origin trial on June 9. When to reach for each, the two API patterns, and a decision table worth bookmarking.