News✳
The releases, research and policy decisions shaping AI around the world. Original context, with sources you can follow.

Gemini Omni 1.2 Flash: longer video, lower pricing and the limits to check
Google’s November 5 video release brings 15-second clips and conversational revisions. The Cloud specifications and Vids rollout give teams a more precise picture of what is available.

Muse Spaces turns a prompt into a personal workspace inside Meta’s AI app
Meta’s November 6 update adds custom experiences inside Muse. The useful question is whether a space can reduce the work of keeping a personal tool current.

Inkling 2 Small arrives on Tinker, with downloadable weights still planned
Thinking Machines Lab’s November 5 release adds a new multimodal model for inference and fine-tuning. Its launch documents separate current Tinker access from a planned Apache 2.0 weight release.

BioBasicsBench asks whether AI can read the evidence in the lab
OpenAI’s November 6 benchmark targets routine experimental analysis. Its results offer a useful lesson in reading AI scores: partial credit and reliable decisions are different measures.

Images 2.6 adds GIFs and new editing controls: what changes for creators
OpenAI’s November 4 release pairs faster image generation with motion and more control over revisions. Developers also get separate settings for scene planning and rendering quality.

Sarvam Prism: a decision model that scores choices instead of writing replies
Sarvam’s November 3 release targets routing, triage and evaluation. Its API charges for input tokens, while the open-weight version has substantial GPU requirements.

LongCat-2.5 reaches the API: coding-agent support, limits and introductory pricing
The November 2 API release targets visual coding and complex project work. Its integration options and limited-time token rates give developers a concrete starting point for evaluation.

Gemini adds in-chat file editing: PDFs on mobile, documents and slides on the web
Google’s October 29 update brings manual and prompted revisions into the conversation. Here is where each format works and what to check before exporting.

OpenAI backs a shared AI-training signal: what website owners should know
The October 30 commitment could simplify publishers’ training preferences. Implementation is still ahead, and the proposal is separate from ordinary crawler rules.

Pika H3 Flex: what creators get locally, and what stays in the cloud
Pika’s October 29 video release puts iteration first. The key buying decision is which tools come with the downloadable model and which require its hosted service.

Gemini 4 Carbon: limited access, launch pricing and the evidence behind its security claims
Google’s October 28 announcement targets cybersecurity and sustained engineering work. Access starts with selected defenders, while the evaluation report adds context to the headline scores.

Rime launches Prism: a voice AI that lets applications check replies before speech
The October 28 release combines direct audio understanding with an inspectable text reply. English-only support, integration work and beta fine-tuning still shape where it fits.