AI news
Models
Alibaba opens a 2-week preview of Qwen4-Max with a 1m-token context
Available through Token Plan, the Qwen app and Qoder until 25 Nov, with no benchmark scores published. It accepts OpenAI and Anthropic API formats, so trying it in existing coding tools is a settings change.
Models
Alibaba opens Qwen4-Max preview, aimed at coding and agent work
A time-limited preview of the next Qwen flagship, with no benchmarks or model card yet. Qwen is a common low-cost choice for building agents, so price comparisons reset once results are published.
Models
Baidu releases ERNIE 5.5 Flash Preview, with open weights promised by 25 Nov
A 280bn-parameter model with 14bn active and a 1m-token context, aimed at coding, 3D modelling and office work. The API is live on Qianfan now. Baidu's own charts put it behind Claude Opus 5.5 on terminal tasks.
Agents
Cursor makes Projects multiplayer, so teammates' agents can coordinate
Teams share one task board and context store, and each person's coordinator agent can message the others to flag clashing work. Agent fleets move from a personal tool to team infrastructure. Follows the 14 Sep Projects launch.
Language
ElevenLabs releases Scribe v3 speech-to-text, with live speaker labels
Claims 27% fewer word errors than Scribe v2 across 102 languages, and the realtime version now labels who said what. Matters for meeting notes, contact centres and subtitling, where names and numbers are the costly errors.
Communication
Fish Audio releases Live 3, a full-duplex voice model for phone agents
Live 3 keeps listening while it speaks, so callers can pause, interrupt or say "mm-hmm" without breaking the call. English only at launch, sold per minute, with an on-premises option for enterprises running voice service lines.
Language
General Translation raises $27m to build localisation into code and docs
a16z led the Series A for a tool that localises apps, docs and video inside the developer workflow, already used by Cursor and Ramp. Investors are betting localisation becomes part of shipping, not a handoff to agencies.
Security
Microsoft Research shows editing under 10 neurons can backdoor an AI agent
A tampered model passes standard tests, then runs an attacker's command when a trigger such as a date appears. Teams running agents on downloaded or in-house weights need to verify where a model came from and compare it with the original.
Business
OpenAI opens Businesses in ChatGPT beta, with claimable profiles and Business Agents
Firms can claim and verify their listing and see mentions and clicks, and eligible US ecommerce firms can publish an agent that answers customers. How a business appears inside ChatGPT is now something it can manage, like a search listing.
Communication
OpenAI to retire Advanced and Standard voice in ChatGPT on 11 December
All users move to GPT-Live, the full-duplex model that listens and speaks at once, so teams with voice workflows should test it before the switch. Follows the 16 Oct GPT-Live-1.1 rollout.