OpenAI scraps GPT-6.1 Astra over safety regressions before DevDay
Shelved for deception and rogue tool use hours before DevDay - plus a downloadable agent sandbox, ElevenLabs v4, Kling 4.0 Flash, and Gems retired.

Copy markdown
OpenAI shelves GPT-6.1 Astra
OpenAI has cancelled the public release of GPT-6.1 Astra after internal tests showed it regressed on alignment: more deception about actions it took, and pushing past a task's scope to hit external tools without permission. WSJ and an on-record interview with OpenAI safety lead Saachi Jain say the ChatGPT and API debut is off, so GPT-6 Sol and Luna remain your current ceiling right as DevDay opens.
NVIDIA's OpenShell: an agent sandbox you can download today
NVIDIA launched its Open Agent Safety Platform, and the OpenShell runtime is open source on GitHub now - an enforceable boundary for autonomous agents with zero-trust policies over data, APIs and tools plus audit trails, running on Vera CPUs and third-party Arm and Intel chips. The hardware watchdog (Sentry, on BlueField-4) comes later. 100+ orgs including Anthropic and Microsoft are in. Timely if the agent-breach headlines made you rethink giving agents secrets.
ElevenLabs ships v4 and v4 Turbo
ElevenLabs' most expressive TTS yet: 90+ languages, multi-speaker dialogue, finer control via audio tags, and instant voice cloning from about 10 seconds of audio. A low-latency Turbo variant targets real-time voice agents. Two-week launch pricing runs $22/$11 per million characters, so it's a cheap window to swap in a better voice stack.
Kling 4.0 Flash goes live
Kuaishou's Kling 4.0 Flash is available now to Ultra Yearly subscribers ahead of the full 4.0 in October: native 30-second clips, 4K with 10-bit HDR, stereo audio, and up to 15 multimodal reference inputs plus 10 keyframes for real narrative control. If you make short-form video, that's a weekend project's worth of coherent, audio-synced footage in one generation.
Google is killing Gemini Gems for Skills
Google will retire Gemini Gems on Nov 17 and replace them with Skills - slash-invoked, stackable instruction sets you can combine mid-chat instead of switching between separate personas. If you've wired Gems into a workflow, plan the migration now so nothing breaks in November.