Vercel AI Gateway adds Kimi K3.1 Pro on zero-data-retention routes
Production access to the near-Opus open model with auto-failover — plus Datasette's agent now builds HTML apps, and GPT-Image-2.6 hits the API.

Copy markdown
Kimi K3.1 Pro, now behind one ZDR endpoint
Vercel's AI Gateway added Zero Data Retention providers for Kimi K3.1 Pro and the faster K3.1 Pro Highspeed, with automatic failover and per-request traces. You can drop the just-shipped 2.8T-param MoE open model into production without self-hosting vLLM or routing prompts through a data-retaining backend.
Datasette's agent now ships custom HTML apps
datasette-apps 0.3a0 lets the Datasette Agent create, host, and edit custom HTML apps straight from chat — tag them with #, and new app_add_stored_queries() / app_remove_stored_queries() tools wire them to your data, with write queries gated behind explicit approval. A weekend build: point it at a database and have it generate a shareable mini-dashboard.
Update: GPT-Image-2.6 reaches the API and Codex
Beyond the ChatGPT GIF rollout, OpenAI brought GPT-Image-2.6 to the API and Codex as two callable models — Flare (~6s per image) and Sunburst (SVG export plus steadier edit-to-edit consistency) — generating about 50% faster than before. You can now wire the new image model directly into your own apps and agents.
Datasette patches a quiet model-allowlist leak
datasette-llm 0.1a10 fixes a bug where the agent wasn't restricted to the models listed in its config — the agent.datasette.io demo had every installed model exposed. Upgrade if you run Datasette's LLM agent so users can't invoke models (and rack up costs) you never meant to offer.