Verdict: You can now use Claude Fable 5 inside Claude Code to automate a large slice of video post-production — transcribing raw takes, picking keeper shots, matching B-roll, adding captions, and exporting clips — without touching a conventional timeline editor. But there is a catch. Anthropic released Fable 5 on June 9, 2026, then suspended access on June 12, 2026, following a US government export-control directive. That means the components are real and the workflow is buildable, yet the exact model may not be available when you read this. Plan around Fable 5 if it returns, or swap in Claude Opus 4.8, GLM 5.2, or another long-context coding model while you wait.
Last verified: 2026-06-18 · Best for: creators who produce recurring talking-head, tutorial, or founder-video content · Caveat: Fable 5 access is currently suspended · Good fallback: Claude Opus 4.8 or GLM 5.2 in an agent harness
What this article covers
This guide walks through a code-first, agentic video editing workflow that uses a frontier language model to do the boring parts of post-production:
- Rough-cut raw talking-head footage by transcript.
- Build a searchable B-roll library with auto-generated descriptions.
- Plan visuals, text overlays, and pacing.
- Render with open-source tools (FFmpeg, Remotion) instead of a paid NLE.
I will not sell you magic. The output still needs a human review pass, and it works best when you feed it repeatable formats. But if you publish weekly video, this can replace hours of manual scrubbing, file renaming, and timeline stitching.
If you are new to AI agents, start with our AI for Small Business: The Complete 2026 Guide before tackling a video agent pipeline. If you already run Hermes Agent as your personal agent OS, the same coordination patterns apply here.
Why agentic video editing is suddenly practical
Three shifts made this workflow possible in 2026:
- Long-horizon coding agents. Claude Fable 5 is Anthropic's first public "Mythos-class" model. Anthropic's launch post says it is built for "days-long, complex, and asynchronous tasks" and can "plan across stages, delegate to sub-agents, and check its own work" when run inside Claude Code or Claude Managed Agents.[^anthropic-fable-launch]
- A 1-million-token context window. Fable 5 can keep an entire project's transcript, asset descriptions, and cut notes in working memory. That matters because video projects accumulate a lot of text metadata.[^anthropic-fable-page]
- Open rendering stack. FFmpeg has always been scriptable; Remotion lets you describe motion graphics and edits in React and render them to MP4. Combined, they give an agent something it can actually execute and verify.
Anthropic priced Fable 5 at $10 per million input tokens and $50 per million output tokens — about double Claude Opus 4.8 — so this is not a cheap chatbot. It is a premium engine for workflows where manual time costs more than tokens.[^anthropic-fable-page]
If you are comparing frontier models for this kind of agentic work, see our earlier comparison of OpenRouter Fusion vs Fable 5 for a multi-model alternative that can match or beat Fable 5 on deep-research tasks at lower cost.
The workflow in five stages
Stage 1: Capture and transcribe raw footage to the word
The agent cannot edit what it cannot read. The first step is to turn every recorded word into a timestamped transcript.
A typical command inside Claude Code looks like:
whisper raw_take.mp4 --model large-v3 --output_format json --word_timestamps True
OpenAI's Whisper is free, local, and accurate enough for this purpose. It outputs a JSON file where every word carries a start and end time. The agent reads this JSON and sees the raw take as an editable script.
Why this beats a timeline: Instead of scrubbing for the good take, you ask the agent, "Find the last clean reading of each sentence and remove false starts, long pauses, and 'ums'." The agent maps its decisions back to exact timecodes.
Stage 2: Build the rough cut from transcript logic
Once the transcript is loaded, the agent applies a small editing policy you write in plain English:
- Drop false starts and retakes.
- Remove pauses longer than 0.8 seconds unless they are intentional.
- Keep the last clean take of each sentence.
- Preserve natural breaths so the audio does not sound robotic.
- Avoid cutting on waveform spikes that would create harsh transitions.
The agent outputs a final-edit.json file that lists selected segments as [[start, end], ...]. A short Python script then calls FFmpeg to concatenate those segments:
ffmpeg -f concat -safe 0 -i segments.txt -c copy rough_cut.mp4
This mirrors what a human editor does on a first pass, but it runs in seconds and scales across hours of footage.
Stage 3: Index your B-roll so the agent can "see" it
For the final edit to feel professional, you need visuals that match what the speaker is saying. The trick is making your B-roll searchable by meaning, not filename.
The workflow is:
- Drop new camera/screen footage into a folder.
- Extract one frame per second from each clip.
- Ask the vision-capable model to describe the frames and note the timestamp.
- Save the descriptions in a plain-text or JSON library.
A good description entry looks like:
{
"file": "broll_001.mp4",
"start": 12.0,
"end": 18.0,
"description": "close-up of hands typing on a laptop keyboard in a bright home office",
"mood": "work, detail, grind"
}
Over time you build a mood-tagged library: detail, freedom, grind, machine, people, place, product. When the agent needs a shot to illustrate "I built the whole thing," it searches for clips tagged with grind or machine and picks the one whose timestamp best matches the narration.
Stage 4: Plan the full edit like an assistant editor
Now the agent has everything it needs to plan the final video:
- the rough-cut transcript,
- the searchable B-roll library,
- your brand fonts, colors, and lower-third style,
- a target format (9:16 short, 16:9 long-form, etc.).
It writes an edit plan that maps each sentence to a visual element:
00:04.120 - 00:08.300: talking head
00:08.300 - 00:12.500: B-roll broll_001.mp4 [12.0-18.0] "laptop hands"
00:12.500 - 00:14.800: text callout: "Build a business that doesn't need you"
The first render can be low quality — a preview MP4 generated with FFmpeg — so you can watch the pacing without burning credits or compute on a full render.
Stage 5: Full render, feedback loop, and final export
After the preview, you give feedback in natural language:
"The clip at 00:08 feels generic. Replace it with something that shows movement, not just a laptop."
The agent:
- Takes a screenshot of the preview at the flagged timestamp.
- Reads the screenshot to confirm the problem.
- Searches the B-roll library for a better match.
- Updates the edit plan and re-renders.
For motion graphics, captions, and lower thirds, the agent can generate React components in Remotion and render them into the final MP4. Remotion's Claude Code integration lets you add the official skill with npx skills add remotion.[^remotion-claude-docs]
If you want a cheaper or fully open-weight planning engine, GLM 5.2's 1 million token context can run the same long-horizon planning loop inside a local or cloud agent harness.
Tool stack and real costs (2026)
| Layer | Tool | What it does | Cost |
|---|---|---|---|
| Agent harness | Claude Code or Claude Managed Agents | Plans, executes, and verifies across sessions | Claude Pro/Max/Team or API tokens |
| Frontier model | Claude Fable 5 | Long-horizon reasoning and multi-stage planning | $10/M input, $50/M output when available; fallback to Opus 4.8 at half price |
| Transcription | OpenAI Whisper (local) | Word-level timestamps | Free if run locally; OpenAI API $0.006/min for large-v3 |
| Cutting/encoding | FFmpeg | Frame-accurate trim, concat, preview render | Free |
| Motion graphics | Remotion | React-to-video captions, callouts, transitions | Free for individuals; team plans from $19/seat/month |
| B-roll search | CLIP-style embedding or vision model | Match clip descriptions to narration | Varies; small open models are free |
| Final polish | Descript / CapCut / DaVinci Resolve | Human review pass, color, sound mix | CapCut free-$19.99/mo; Descript from $16/mo; DaVinci Resolve free |
Source notes: Anthropic Fable 5 pricing from Anthropic's Fable page.[^anthropic-fable-page] Remotion skill install and docs from remotion.dev.[^remotion-claude-docs] CapCut and Descript pricing from official pages and verified third-party trackers as of June 2026.[^descript-pricing] [^capcut-pricing]
Important: Fable 5 is currently unavailable
On June 12, 2026, Anthropic announced it was suspending access to Fable 5 and Mythos 5 after receiving a US government export-control directive. Anthropic says it disagrees with the action and is working to restore access, but there is no public timeline.[^anthropic-access-statement]
What to do now:
- If you already have Fable 5 access, expect it to be paused.
- If you are building this workflow, write it against Claude Code with model-agnostic agent calls so you can swap models.
- Use Claude Opus 4.8, GLM 5.2, or GPT-5.5 as the reasoning engine for now. The workflow is mostly about long context, tool use, and iteration — not Fable 5 alone.
What this means for you
For a small business or solo creator, this workflow is not a license to fire your editor. It is a way to eliminate the first 70% of repetitive post-production so the human editor (or you) can focus on taste, pacing, and final polish.
Best fit:
- Weekly founder videos, tutorials, or explainers with repeatable formats.
- Podcast or interview clips repurposed into social shorts.
- Teams that already use Claude Code for code and want to reuse the same harness for media.
Worst fit:
- One-off cinematic projects with heavy color grading and sound design.
- Workflows where every frame is custom and non-repeatable.
- Teams that do not have someone who can review the agent's output before publish.
FAQ
Q: Do I need to know how to code to use this workflow?
A: You need some comfort with the terminal and JSON, but the agent writes most of the code. Claude Code runs FFmpeg and Remotion commands for you, and the skills system means you can install pre-written instructions. You still review the output like a producer, not a programmer.
Q: Can this really replace Adobe Premiere or DaVinci Resolve?
A: Not for everything. It can replace the repetitive parts: rough cuts from transcripts, B-roll matching, captions, and simple motion graphics. It cannot yet match the color science, audio mixing, or fine-grained keyframe control of a professional NLE. Most users will export a first cut from the agent and finish in Resolve, Premiere, or CapCut.
Q: How much does it cost to edit one video this way?
A: If you run Whisper locally and use a mid-tier Claude subscription, the marginal cost per video is mostly your time and model tokens. A 10-minute talking-head project might consume a few hundred thousand to a few million output tokens. At Fable 5 prices, that could be tens of dollars per video once access returns. At Opus 4.8 or GLM 5.2 prices, it drops sharply. Treat it as a trade-off: token cost versus editor hours saved.
Q: What if Claude Fable 5 does not come back soon?
A: Build the workflow model-agnostic. The same pipeline works with Claude Opus 4.8, GLM 5.2, or GPT-5.5 as the planner. Fable 5's advantage is stamina on very long and complex tasks, not a monopoly on the idea.
Q: Is the output good enough to publish straight to YouTube or LinkedIn?
A: For social-native talking-head clips, yes — after a human review. For long-form YouTube documentaries or brand commercials, no. Use the agent for the rough cut and platform versions, then send the final pass to a human editor or polish it yourself.
Q: Where do I store the B-roll library so the agent can access it?
A: A cloud folder works (Google Drive, Dropbox, or an S3 bucket), but the agent needs a local or mounted copy to process frames. A practical setup is: store originals in the cloud, sync the current project's assets to the machine running Claude Code, and keep a description index file in the project folder.
Discussion
0 comments