Modal open-sources Quail: run LLMs over your data at 10x vLLM
Quail treats LLM-over-SQL as a database problem, hitting 1B tokens/min/GPU. Plus: Claude Code model pinning, Gemini's live avatar, an OpenAI 'o' leak.

Copy markdown
Quail: LLM-over-SQL at database speed
Modal open-sourced Quail, an inference engine that treats AI-SQL — SQL with LLM filters and joins — as a query-planning problem. It clears 1B tokens/min per H100 on multi-join queries, roughly 10x the vLLM baseline (1.84x geomean across its suite), using KV-aware join ordering and an 8-token vocabulary for boolean classification. It's `pip install quail-engine==0.1.0` and runs on Modal today — which makes batch-classifying or filtering millions of rows with a model finally cheap enough to ship.
Claude Code can now pin models and block new ones
Claude Code v2.1.283 adds two managed settings — `deniedModels` and `availableModelsMatch: "exact"` — so teams can freeze the model list and stop a fresh release from silently swapping in mid-project. A new `/doctor prompt-audit` scans your CLAUDE.md, skills, agents and commands for stale model references. If a surprise model rev has ever broken a prompt or blown a budget, this is the off switch.
Gemini 3.8 Live avatar hits GA — voice agents over the API
Google made Gemini 3.8 Live with Live Avatar generally available: speech-to-speech with natural interruption recovery, live camera and screen input, tool calls mid-conversation, and 97-language auto-detection through the Live API and Agent Development Kit. Individual builders can wire up real-time voice and video agents now; the lip-synced custom avatars stay enterprise-allowlist for the moment.
Report: OpenAI's always-on 'o' agent leaks before DevDay
A ChatGPT Pro upgrade screen reportedly surfaced "o, your always-on assistant" — a persistent agent with its own identity that keeps working across chats after you close the app — hinting at what OpenAI may unveil at DevDay on Sept 29. It's unconfirmed by OpenAI, but several outlets caught the same screen. Worth watching if you build on the Assistants or Agents stack.