AI news this week: Qwen puts Qwen-Image-2.1 weights on the public internet, Runway adds a frame-rate model and one credit pool

This week, 16-22 September 2026, three things moved for people who make things with these tools. Qwen published the weights of Qwen-Image-2.1, a single image model that generates and edits, on 20 September. Runway put out two announcements in two days: a dedicated frame-rate model, and one credit pool across its web app and its developer platform. And Reve told its users that image generation there is being switched off, with a download deadline. Nine items follow, grouped by what you make with them. Every one was opened at the vendor's own page and dated from the text on it; where a figure is the vendor's own, it says so.

Video

Runway ships Enhance Frame Rate

Runway announced Enhance Frame Rate on 17 September, a model that converts any video - whether or not it was made in Runway - to 25, 30, 48, 60 or 120 fps, including the 59.94 fps rate. The point is the delivery spec: a clip that arrives at the wrong frame rate needs a manual conform pass before it can sit in a timeline, and this removes that step. Runway says the model runs up to seven times faster and three times cheaper than other interpolation models, with higher output quality; that is the company's own comparison and we have not reproduced it. It joins Runway Ruby, the SDR-to-HDR upscaler Runway introduced earlier this month, so a generated clip can now be conformed and graded without leaving the platform.

Runway puts the web app and Runway Dev on one credit pool

One day later, on 18 September, Runway changed how its own money works. Purchased credits are now a single pool, usable across both the web app and Runway Dev, bought on one invoice, with a revamped workspace analytics view so admins can track and move credits across both kinds of workspace. The enterprise offers are the vendor's words: new enterprises that sign up by 31 December 2026 receive 10% more credits, and existing enterprises get 5,000 complimentary credits plus one free Applied AI Architect day, which Runway values at over $6,000. Read them as an offer, not a measured price.

Higgsfield opens a pay-as-you-go API

Higgsfield published the Higgsfield API on 16 September: a self-serve interface for developers with no subscription, where you top up a balance in US dollars, pay per generation at published rates, and spending stops automatically at zero. The company says it exposes more than 50 current video and image generation models - the same catalogue you meet on the site, Seedance 2.5, Kling 3.0 and Flux 2 among them. Two details decide whether it suits you: the balance is separate from any Higgsfield plan, so a subscription neither grants access nor changes the rates; and funds expire one year after they are added, so a large top-up is not a permanent credit.

Audio

ElevenLabs rebuilds Studio inside ElevenCreative

ElevenLabs published Studio 4.0 on 21 September, a major upgrade to the video editor inside ElevenCreative. The idea is that you stop carrying footage between four tabs: the shot, the voiceover, the music and the sound effects are generated inside the project, on the timeline, and you cut, caption and export there. Voiceovers draw on over 10,000 voices in 32+ languages, the company says, and generation can be repeated in place - a shot that is nearly right is refined where it lies, without exporting, re-uploading and rebuilding the edit around a new file. The rebuilt timeline adds frame-level zoom for exact cuts and clip snapping. If your work is mostly the audio and you only export the video, this is the week's most relevant update in that corner of the market.

Descript ships generated music, a caption rebuild and per-second smoothing

Descript's release roundup dated 17 September on its changelog is the biggest drop of audio features this week. Descript's AI can now generate background music and sound effects from a description and drop them on the timeline at the length you ask for - a whoosh, a doorbell, a knock - which removes the stock-music subscription from the workflow (it needs version 145.0.15 or newer). Captions were rebuilt: 18 new presets and 10 new text animations, 15 in total, each with separate in and out motion. The voice picker now draws from the ElevenLabs voice library instead of Descript's own short set. And jump-cut smoothing has moved from a flat charge per cut to a rate per second, which changes what a rough edit costs.

Image

Qwen releases Qwen-Image-2.1 with open weights

Qwen put Qwen-Image-2.1 on the public internet this week: weights on Hugging Face, announced by the project's X account at 06:06 UTC on 20 September. It is one model for both text-to-image generation and editing. The card reads a visual generation component of just 7B parameters in 32 single-stream DiT layers, native transparent (RGBA) output and editing that takes up to 10 reference images, with local edits marked by circles, painted annotations or separate masks, and identity preserved for people and products. One caveat before you build on it: the model card records the licence as the Qwen Research License Agreement, which is a research licence rather than a permissive open-source one, so check the terms before you ship it inside a paid product. ComfyUI shipped support for it in v0.37.0 the next day, which is the fastest route into a self-hosted pipeline.

Reve switches off image generation

Reve is sunsetting. Its help centre now states that after 27 September 2026 creating new images and videos will no longer be available, and that you have until 31 October 2026 to download what you already made. Plan the extraction: prepare a download once every 24 hours, and note that a prepared link is valid for only two days, so download your library in passes, not one. Krea - which sells a competing generator and therefore is not a disinterested witness - published a migration guide on 18 September saying Reve announced the sunset that day, ended image generation at 11:59 PM Pacific on 27 September, and stopped taking new accounts and paid subscriptions on 17 September. Krea's dates match the vendor's own notice, which is why the item stands; read the commercial parts as advertising.

Harnesses & open source

ComfyUI v0.37.0 reaches for the week's image release

Comfy-Org released v0.37.0 of ComfyUI, published at 07:35 UTC on 21 September in its GitHub release feed. The changelog adds Qwen-image 2.1 support, so the weights above are runnable in the node editor the same week they landed; it also adds MoGe 3, Aimdo 0.5.5 with automatic --fast-disk detection, a CFG control on the YuE2 Generate Music node, transparent-background support for the GPT Image 2 partner node, and CUDA-graph and w4a8 work for the Qwen3/3.5/3.8 text encoders. If you maintain a local install, this is the release to pull.

Codex CLI 0.155.0 adds /voice; 0.155.1 puts the default back

OpenAI's ChatGPT and Codex changelog dated 17 September records Codex CLI 0.155.0, which added experimental voice conversations with live transcripts and microphone controls, off by default and enabled through /experimental, along with live reasoning summaries in the status row. The following day brought 0.155.1, which restored none as the default for terminal reasoning summaries - a fix, in the project's words, for requests rejected by providers that do not support them (issue #46467). If you run the CLI against a third-party provider, take 0.155.1 or later rather than the 0.155.0 everyone shipped on the 17th.

Items were checked against their primary sources on 22 September 2026. Dates above are the dates printed on the vendor page, not the date a newsletter carried them. Where a number is the vendor's own - the seven-times-faster claim, the credit offers, the model count - it is attributed in the text and has not been reproduced by us. Nothing here is advice on where to spend your subscription money.

We use AI tools in our research and drafting on this site. Every judgement and every recommendation is a person's.