OpenAI Ships Images 2.6, ElevenLabs Hits SoundCloud, Inkling 2 Small Lands

OpenAI released Images 2.6 this week, and the headline addition for creators is native GIF generation alongside faster still-image output and more precise inpainting controls.

The GIF support alone changes what you can do inside a single workflow. Short looping visuals for social posts, product demos, or sticker packs no longer require a separate tool. The new API controls also give developers finer command over output style and edit regions, which matters if you're building image pipelines rather than one-off generations.

The update rolls out across ChatGPT, Codex, and the API, so wherever you already work with OpenAI's image tools, the improvements show up without any migration. Start by testing the inpainting precision on a complex product photo where previous versions left visible seams.

ElevenLabs has partnered with SoundCloud so you can publish AI-generated music directly from ElevenLabs to your SoundCloud profile, as The Verge reports.

This is the distribution gap closing in real time. Generating a track was already fast; getting it in front of listeners required exporting, formatting, and uploading manually. A direct publish path removes that friction entirely and puts AI music creators on the same footing as anyone else uploading to the platform.

If you're already using ElevenLabs for voice or sound design, the SoundCloud integration is worth connecting now even if you don't plan to release tracks immediately. Having the pipeline set up means you can ship the moment something sounds right.

Thinking Machines Lab released Inkling 2 Small, a Mixture-of-Experts model with 594B total parameters but only 17B active at inference time, which is the number that actually determines how much compute you need to run it.

It supports image inputs and variable effort settings, meaning you can dial inference compute up or down depending on the task. The model scores 31 on the Artificial Analysis Intelligence Index and runs on the Tinker platform. Apache 2.0 weights are planned, which means local deployment without licensing restrictions.

For creators running local setups, the active-parameter count is what to focus on. 17B active on a MoE architecture can deliver quality well above what a dense 17B model would produce. Watch for the Apache 2.0 weight release and test it on your longest, most complex generation tasks first.

Perplexity launched Teammates, a set of always-on AI agents that take on defined roles inside a shared workspace, as The Verge covered.

Each Teammate can hold a specific function, like research, writing, or summarization, and pulls context from projects, memory, and tools you're already using inside Perplexity. The always-on framing means they're not waiting to be prompted; they can surface relevant information as work evolves.

If your team already uses Perplexity for research, Teammates is worth piloting on a recurring workflow first, something like weekly competitive research or brief generation, where the persistent memory and role assignment will show the most obvious return.

Google updated its video generation model with Gemini Omni 1.2 Flash, pushing native video generation and editing up to 15 seconds with improved frame consistency across the clip.

Consistency has been the persistent weak point in AI video, where subjects drift and scenes lose coherence past the five-second mark. Google says this version addresses that directly, which makes the 15-second ceiling more usable in practice than it sounds on paper.

Test it on scenes that require a subject to move across the frame or hold a recognizable form for the full duration. That's where previous Flash versions broke down, and it's the fastest way to see whether the consistency improvement is real for your use case.

The through-line this week is friction removal. OpenAI's GIF support and ElevenLabs' SoundCloud integration both shrink the gap between making something and getting it out into the world. Perplexity's Teammates and Gemini Omni 1.2 Flash extend how long an AI can hold a coherent task or a coherent frame, which are related problems solved in different domains.

Inkling 2 Small is the one to bookmark for later. Apache 2.0 weights on a 594B MoE with 17B active parameters could be the most capable locally-runnable model yet when the weights drop. Set a reminder, because the window between release and the community figuring out optimal quantization settings is when you want to be testing.