Schematron V2: cheap HTML-to-JSON extraction, now on OpenRouter
Two 3B models turn any page into schema-locked JSON for pennies, beating GPT-5.4 Nano. Also: Claude Code's git-read nag fixed; Copilot joins the Vercel AI SDK.
Copy markdown
Schema-locked JSON from any page
Inference.net's Schematron V2 Turbo (3B) and Small are now on OpenRouter: feed messy HTML plus a JSON schema in response_format and get valid structured data back every call, no prompt-coaxing. Both run a 128K context and beat DeepSeek V3.2 and GPT-5.4 Nano on extraction quality, with Turbo at $0.03/M in and $0.15/M out — cheap enough to point at a whole category of sites and build a clean dataset or price monitor without babysitting parse errors.
Claude Code quits nagging on git reads
Claude Code 2.1.270 fixes a 2.1.269 regression where read-only git commands (status, log, diff) started asking for permission again after a session had been running a while. If you've spent the week re-approving harmless git reads, updating clears it.
Copilot lands in the Vercel AI SDK
Vercel shipped @ai-sdk/harness-github-copilot, exposing GitHub Copilot through the AI SDK's HarnessAgent interface. You can now drive Copilot from the same code that runs your other agents — handy if your org already pays for Copilot seats and wants them scriptable alongside everything else.