Everything we've published, newest first.
We are huge fans of Tim Simmons' Theoretically Media, and this week he rebuilt his 2025 short film by pointing GPT-6 in Codex at Runway's MCP, then finished the cut in DaVinci Resolve without touching the mouse. Watch it, then subscribe to his channel.
Lightricks' open trainer turns your clips into a small adapter that teaches the 22B video model a style, a face or a camera move. Here is the whole run.
The studio behind Tribeca's Dreams of Violets says its second feature ran on Kling and took months of part-time work, and reviewers have not been kind.
Musk's single-word quote of someone else's clip drew 6.8 million views, while xAI's own price list puts its current video model at 8 cents a second.
Alibaba's Wan 3.0 doubles its predecessor's single-pass clip length to 30 seconds, but its per-second price beats or trails Wan 2.7 depending on which platform bills it.
The Ninth Circuit held that Copilot and OpenAI's Codex generating new code without copyright notices is not "removing" them, leaving infringement claims as the harder road for AI plaintiffs.
MiniMax H3 through a speed-tuned reseller undercuts everything at $0.02 per output second, while ByteDance's Seedance 2.5 is the priciest name that still publishes a real rate.
A launch discount briefly makes Alibaba's Wan 3.0 cheaper per second than its own predecessor Wan 2.7, and that discount, like Pruna AI's, ends September 24.
OpenAI's Videos API and every sora-2 model ID go dark September 24, and the replacements price per second, not per token, across Google, Alibaba, fal and Black Forest Labs.
Kuaishou confirmed 15-second multi-shot clips and five-language native audio for Kling 3.0; the widely repeated 4K 60fps and six-shot figures trace to a single marketing release.
HiDream.ai's new omnimodal model can pick its own clip length between 5 and 20 seconds, and the company says it debuted at No. 4 and No. 8 on two video leaderboards.
The change merges credit purchases into one invoice and adds sign-up bonuses for enterprises that commit by December 31, 2026.
A composer in 2112 hears the machine speak in a photoreal AI short whose night-city craft is a cut above the beat, and whose ending is still unproven.
Eight competitions publish about $6 million in creator prizes, and the largest settles its field at a live Los Angeles final on September 25.
In a talk recorded six weeks before Jev launched, TypeSafe's founder argued AI was trained to please people, which is exactly why it cannot be trusted to decide.
TypeSafe's Jev answers typed questions with calibrated probabilities instead of writing text, and output tokens are billed at zero.
Meta withdrew its Muse Image likeness feature after three days, then built a labeling rule that spares honest AI personas and mislabels humans instead.
Parent company Pocket Entertainment reportedly seeks $100 million to $120 million at up to a $2 billion valuation, and says AI made production about 80 times cheaper.
The open weights logged 24.8 million downloads, but the stock now sits 77% below its March peak, and nothing proves the giveaway caused the rally.
Nine entries are ranked on YouTube view counts read on September 19, 2026; No. 7 is our own short film, included as a pick rather than on the numbers.
