Creative AI Briefing November 6, 2026 morning edition 12 min read

This week AI showed up inside software creators already open every day, often at no extra cost, so a free stem splitter and a free 15-second video generator are now part of the baseline kit and paid standalone tools have to beat what is already in the app.

Gemini Omni 1.2 Flash Google Vids Pro Tools 2026.11 Kapwing MCP Phonon TTS sprite-gen MiniMax H3 video-gen audio-gen ai-editing creative-workflows frontier-models

Creative AI Briefing: Friday, November 6, 2026

Open Google Vids today and you can generate a 15-second video clip with sound, at no additional cost on a Google account, using the Gemini Omni 1.2 Flash model Google released yesterday. Open Pro Tools after updating to 2026.11 and you can split a mixed song into six stems on your own machine, even in the free Intro version. Neither feature asks you to learn a new app or buy a new subscription. That is the pattern of the week: AI is moving into the tools creators already have open, and it is arriving bundled rather than billed separately. The question for every paid standalone AI tool just got harder: what do you do that the app on my screen does not already do for free?

New models

Gemini Omni 1.2 Flash makes 15-second clips and lets you edit them by talking. Google DeepMind's announcement (November 5, 17:00 UTC) stretches native clip length to 15 seconds, "up from 10 seconds with Omni 1.1," set in one-second steps from 3 to 15. You draft at 360p or 720p, then upscale to 1080p or 4K, and you can keep editing a clip over several turns of conversation. References can include images, short motion clips "up to three seconds" and audio cues, and Google says speech sounds more natural with "clearer lip-sync" across several speakers. The API price is "$0.07 per second for 720p video," which Google calls a 30% drop from Omni 1.1, so a full 15-second 720p draft runs about $1.05. Where it runs: the Gemini API through Google AI Studio, Google Cloud, Flow, Google Ads, the Gemini app ("rolling out to Google AI Ultra subscribers") and Google Vids, where 15-second clips come "at no additional cost" with a Google account or Workspace subscription. Every output carries a SynthID watermark. Read the Vids fine print before you plan on that free tier for client work: Google's Vids help page says "Generated images and videos are for use only within Vids" and caps most users at 50 generated videos a month (that page still describes 3 to 10 second clips, so it may predate this update). The honest catch is Google's own: work continues on "fast sports action, rapid script changes and complex physics." A second catch for API users: at our check, the Gemini API pricing page and Omni docs still listed only gemini-omni-1.1-flash (about $0.10 per second at 720p, no free tier), with no upscale price shown. Confirm the 1.2 model name and price in AI Studio before you budget a project around $0.07.

Video

Kapwing now edits video from inside Claude and ChatGPT. Kapwing's MCP page (MCP is the standard way to plug an app into an AI assistant) says the connector works in Claude Web, Claude Desktop, Claude Code, Cowork and ChatGPT. You can upload footage, edit, generate video from a prompt, add and translate subtitles, dub, sync B-roll and make voice clones by asking in the chat. There is "no separate MCP fee": it spends the credits of your existing plan and works "on all Kapwing plans, including free accounts." The page gives no launch date (it surfaced on Show HN on November 5) and warns that "you may deplete credits faster than expected when using the MCP." Kapwing "does not at this time have an API or SDK for developers," so this is the only programmatic door.

Replicate published a plain buyer's guide to MiniMax H3. The November 3 post lays out three tiers by price per second of output: H3 Lite at $0.025 (480p) or $0.04 (768p), H3 Highspeed at $0.06 (768p only), and full H3 at $0.08 (768p) or $0.13 (2K). All three make 4 to 15 second clips from text, a first frame (with optional last frame) or references of up to 9 images, 3 videos and 3 audio files. The most useful line is a warning: "Writing '2K' in the prompt does nothing. Set the resolution input." The workflow section below turns this into a drafting ladder.

Aireel.tv turns a landing page into a promo film. Shown on Hacker News November 5, it promises script, motion design, voiceover and score from one pasted URL, 15 to 90 seconds, with five rounds of notes. A free 15-second ad carries a small watermark and "may be featured as an example on our site." Paid per-film prices were not visible on the page at our check, and director's-cut films from $79 plus monthly Studio ($299) and Agency ($799) plans are all marked coming soon. Read the terms before you put a client's page into it.

DaVinci Resolve 21.2 is out. Per RedShark News (November 5), the free update adds Gen 6 LUTs and MacBook Pro touch support, and Resolve Studio gets "improved AI assistant color and Fusion capabilities." Studio is $236 (20% off) until November 20.

Audio and music

Pro Tools now splits stems for free, on your machine. Avid's 2026.11 update, reported by RedShark News (November 5) and MusicTech, adds an AI Stem Separator that pulls up to six stems (vocals, drums, bass, guitar, keys, other) locally with no cloud upload. It is "included at no additional cost with all versions of Pro Tools, including the free Pro Tools Intro." It reaches Artist, Studio and Ultimate subscribers, perpetual owners with an active upgrade plan, and Intro users. For an editor who needs the music out from under a voice, or a producer sampling an old bounce, the cost of a first pass just went to zero. Avid's own release notes were not read this run, so the separation quality and any track-length limits are unverified here. The paid alternative covered here yesterday, iZotope's Photon, now has to sell its denoising and finer drum splits against a free baseline (RedShark, also November 5, lists Photon at $349, or $259 with cyber deals through December 7).

Phonon TTS explains its licensing. Yesterday's briefing covered Phonon running in a browser tab. The Hugging Face deep dive (November 5) adds the terms that matter for paid work: the English Phonon 31m model is Apache 2.0, usable "even in a commercial product," while the five-language Phonon 40m (English, French, German, Spanish, Portuguese) ships under a non-commercial license (the post calls the other models CC BY-NC) unless you pay for one, with "pricing based on your number of users, not how much speech you generate." The models are about 65 MB and 75 MB, ship with more than 100 voices, and run on phones, laptops and in a browser. Benchmark claims (first in German and Spanish intelligibility, 28 ms to first audio) are the authors' own. Try it in the free Space.

Suno posted "Start Songs with MIDI" on its blog on November 4. The post itself would not load for us, so we cannot say which plans or what limits apply. Suno Studio already handles MIDI tracks, per its September 17 release note.

Design and 3D

Research, no download yet: IBM's Palette (November 3) fine-tunes a model on about 1,000 IBM decks so it builds slides "in your design language," with plain-English directives like "[[make this a timeline]]" and a repair step that fixes overflowing text. IBM describes a 20B open-weight model, but gives no license, download link or demo.

NVIDIA's Cosmos 3 world models (Hugging Face, November 5) turn simulator controls and text into photoreal video under the OpenMDW 1.1 license, which allows commercial use. They target driving and robotics and were tested on a GB200 data-center chip. Interesting to watch, not a creator tool yet.

Open and local

The local story today is the free stem splitter inside Pro Tools, which runs on your own computer, plus a run of small open tools for game artists and streamers. The two below that interest creators most both came out of Trendshift's daily list. Star totals come from shields.io.

aldegad/sprite-gen: turns one character drawing into a transparent, game-ready sprite sheet plus a manifest file, through Codex, Grok or OpenAI image calls, and installs as a Codex or Claude skill. Apache 2.0, about 3.8k stars, trending this week (repo).
livliv-dev/player: a roughly 30 KB web player for 2.5D characters with lip sync and webcam face tracking, plus VRM and PNGTuber support. Code Apache 2.0, but most bundled characters are personal, non-commercial use only. Created November 4. GitHub's API showed 308 stars and shields.io 738 at our checks; treat both as rough (repo).
microsoft/VibeVoice: Microsoft's open voice model project, back on Trendshift's daily list. An older project, not a new release (repo).
Sandflow-ai/sand3d: "3D modeling in Python, built for LLMs," so an AI assistant can write a model as code you can edit. Creation date not verified (repo).
wzc520pyfm/gen-ai-pose-image: generates images from pose sketches. Trending on Trendshift; creation date not verified (repo).
gradium/phonon (Space): the free browser demo for Phonon voices, the fastest way to judge whether the English model is good enough for narration (Space).

No new image, video or voice model on Hugging Face had a creation date between November 2 and 6 among the trending lists we checked. ComfyUI's latest release is still v0.43.0 from October 27.

Creative workflows

1. One character drawing to a game-ready sprite sheet, with sprite-gen

The steps. Make a Python 3.11+ virtual environment (python3 -m venv .venv && source .venv/bin/activate), then run pip install -e . inside the cloned repo. Put your character on a plain background as base.png. Run sprite-gen prepare --out-dir run1 --character-id hero --base-image base.png, then sprite-gen gen-set --run-dir run1 --provider codex, then sprite-gen extract --run-dir run1 and sprite-gen compose-atlas --run-dir run1. The optional sprite-gen curation step lets you "pick, nudge, breathe." For moving loops, sprite-gen video-set --base side=still.png --states idle,walk,run,jump,attack --out-dir set/ uses Grok's video model.

How it works. Your base drawing locks the character's identity across every pose. The tool generates the animation rows several at a time, keys out the background, and packs the frames into sprite-sheet-alpha.png with a manifest.json listing frame positions, frame rates and loop flags. It can also export Aseprite JSON for engines such as Phaser and Flame.

Why it is good. It ends at a file a game engine can read, not a folder of pretty pictures, and it makes per-state GIFs and contact sheets "so motion is judged as motion before anything ships."

Where it breaks. The README says walk and run cycles "stay experimental unless motion QA actually passes." The video route needs your own Grok login or a paid xAI key. The left/right facing detector "can be wrong even at high confidence." Codex and Grok use your subscription login first; an API-key call is billed per call, and the tool says so "before the request leaves."

2. Draft cheap, finish expensive, with MiniMax H3's three tiers on Replicate

The steps. Explore ideas and batches on H3 Lite at 480p ($0.025 per second, so about $0.25 for a 10-second test). Once a shot works, move the same prompt and references to H3 Highspeed at 768p ($0.06 per second) for "everyday work and reference-driven iteration." Render only the keeper on full H3 at 2K ($0.13 per second), and set the resolution input, because typing "2K" into the prompt does nothing.

How it works. All three share the same inputs (text, first and last frame, up to 9 image, 3 video and 3 audio references), so a prompt that works in Lite carries forward without rewriting.

Why it is good. Ten exploratory 10-second takes on Lite at 480p cost about $2.50, against $13 for the same ten on full H3 at 2K.

Where it breaks. Replicate says "Pick Lite to save money, not time," so drafts are not faster. A shot that looks right at 480p can change when re-rendered on a different model, so expect a couple of finishing takes.

Worth testing

Gemini Omni 1.2 Flash in Google Vids: free 15-second clips with audio on a Google account. Tradeoff: Google's help page says Vids clips are "for use only within Vids," a SynthID watermark on every clip, Google's own warning about fast action and physics, and API docs that had not caught up at our check.
Pro Tools Intro stem separator: six-stem splits at no cost, offline. Tradeoff: quality untested by us, and you are learning a full DAW to get it.
Kapwing MCP: subtitle, dub and edit by asking Claude or ChatGPT, even on a free plan. Tradeoff: you "may deplete credits faster than expected," and the agent's choices are harder to audit than your own timeline.
Phonon Space: free browser narration. Tradeoff: only the English model is cleared for commercial use without a license.

What actually matters from today's signal

The free tier just moved. A year ago, stem separation was a paid plugin and a 15-second generated clip was a credit purchase. As of this week both come bundled with software millions of people already have: Pro Tools Intro and Google Vids. Kapwing did the same thing from the other side, putting its editor inside the chat window people already use and charging nothing extra for the door. For a working creator, that means the first pass on a lot of jobs costs nothing but time, and the money question moves to the finishing pass.

That squeezes the standalone AI tool. Photon's sales pitch was local, private stem splitting; Avid now gives a version of that away. Paid video generators sell clip length and quality; Google now gives away 15 seconds inside a slideshow app. The products that survive will be the ones that are clearly better at the last 10%, not the ones that merely do the thing.

The counter-signal is control. Bundled AI comes with the bundler's choices: Google's watermark, Google's content rules, Avid's six fixed stem types, Kapwing's credit meter running inside a chat you cannot always audit. Free first passes are a gift. Just keep the finishing work somewhere you can see every decision.

Source access notes: Blocked or unreadable this run: openai.com/news (client error), blog.adobe.com (client error), blog.google DeepMind index (JavaScript only; the Omni 1.2 post itself loaded), midjourney.com/updates (JavaScript only), Suno's "Start Songs with MIDI" post (client error; only the blog index entry was seen), api.github.com for sprite-gen (stars taken from shields.io). The shell in this run had no direct network access, so Hugging Face API, shields.io and Hacker News data were read through a fetch tool instead. Hugging Face trending lists showed no new creative model created November 2 to 6. Hacker News coverage stops at about 17:57 UTC on November 5. Writing section omitted (nothing creator-relevant). An adversarial fact-check pass ran on this draft and caught eight problems, all fixed: two misquoted lines and a misread "four at a time" in the sprite-gen workflow, a wrong provider list for sprite-gen, a paraphrase presented as a Kapwing quote, an unsupported Resolve 21.2 feature, Photon misdated to last week, and Aireel tiers lumped together. It also tightened the Phonon and Palette license wording. Unverified after the pass: livliv's creation date and GitHub API star count, the Show HN timing for Kapwing and Aireel, and the Trendshift placements.