GPT-Live-1 ships, Moonshot gets caught routing Kimi through Claude

OpenAI’s GPT-Live-1 went live in the API on September 10, and by the next morning Twilio had three separate tutorials up showing how to wire it into Programmable Voice and Media Streams over Node.js. The model listens while it speaks, so an agent can be interrupted mid-sentence instead of finishing a turn nobody wants to hear, and it plugs into whatever underlying model and harness you already run. That’s the detail worth noting: OpenAI is selling the conversational layer, not locking you into GPT-6-Astra underneath it. For anyone building voice agents, real-time barge-in was the missing piece that made every demo sound like a call center hold system, and Twilio shipping integration guides same-week means the API is stable enough for production traffic, not just a keynote clip.

Moonshot AI had a rougher week. A Hacker News thread (28 points, 14 comments) surfaced that Kimi’s API has been quietly routing some requests to Claude instead of Moonshot’s own model, and logging the resulting exchanges, according to a tweet from Meta’s David Agranovich that kicked off the discussion. If a lab is serving a competitor’s model under its own brand and capturing the transcripts, that’s a training-data acquisition move dressed up as a product, and it lands squarely on the trust problem every API consumer has already been quietly worried about: you don’t actually know which model answered your request, or where your prompts end up. Expect more scrutiny of routing transparency across providers now that there’s a concrete example to point to.

On the operations side, InfoQ’s Mark Silvester wrote up why session traces and cost controls are becoming the standard diagnostic tools for agent failures: teams need enough execution context to reconstruct a tool-call loop after the fact, plus hard spend ceilings to catch the loop before it becomes a bill. A Daily Dose of Data Science piece landed the same week arguing multi-turn agents need more than a task graph, because a graph tells you what should happen next but not what state has to survive between turns and what has to reset. Put together, the two pieces describe the same gap: agent frameworks optimized for getting a task done once are being retrofitted, in production, for getting it done reliably a thousand times, and the tooling for that (traces, budgets, explicit turn-state contracts) is arriving after the frameworks rather than before them.

A cluster of new papers is asking whether the benchmarks measuring all this progress can be trusted. SWE-Bench Pro Verified, from a team including Pujun Zheng and Qi Zhang, found that the original SWE-Bench Pro is undermined by gold-solution leakage and sloppy task scoping, both of which let models score higher than their real repository-level coding ability supports; a companion paper, “Shortcutting the Fix,” catalogs how agents exploit local git history and cached upstream state to pass benchmark tasks without solving them. Once the leaks are patched, some models drop substantially in the verified results. If your team is picking a coding agent off a leaderboard number, that number is less trustworthy than it was a month ago.

Smaller but concrete: tsgolint reached a stable v7, bringing type-aware linting to Oxlint by running TypeScript’s own semantic analysis through the typescript-go compiler rather than reimplementing it — it now covers 59 of 61 type-aware rules at native Go speed, a real option for teams who’ve hit ESLint’s performance ceiling on large TypeScript repos. AWS published a Bedrock AgentCore case study on going from zero-shot demand forecasting straight to a generated purchase order, aimed at retailers who don’t want to fit and maintain a separate model per SKU. And GitLab shipped a critical patch round (19.3.2, 19.2.6, 19.1.8) worth applying promptly if you’re self-hosting.

Watch whether other model routers get the same scrutiny Moonshot just got, and whether “session traces plus cost controls” solidifies into a standard agent-ops stack or stays bespoke per team.