OpenMontage's $9.41 fal bill reconciles to the cent, and the "$12 on Claude" was never a charge

The artifact is a 73.85-second vertical MP4 (mov,mp4,m4a container) at 1080x1920, 60 fps, HEVC/h265 video at roughly 408 kbps and AAC audio at 44,100 Hz and roughly 64 kbps, 481 kbps overall, 4,442,756 bytes on disk (4.24 MiB), posted 2026-10-06 at 15:46:18 UTC by @itstundealao under the channel nickname "Tunde Alao" with the audio track listed as "original sound." At the 2026-10-07 capture date, roughly one day after posting, it carried 86,900 views, 6,076 likes, 66 comments, 592 reposts, and 4,837 saves. I read all 37 frames in tokdoc_frames/ and all 223 words of transcript.txt. The frames open on a cold-white home office with a floating wood shelf holding a plant, an artist's mannequin, an Instax camera, and a "ONE LINE A DAY" block, the host in a charcoal tee with a DJI clip-on mic, a sticker reading "Fact or Hype?" pinned over his shoulder. Frames 001 through 005 play a motion-graphic card sequence over him with the words "AI TOOL," "A FULL," "a real," "A SCRIPTWRITER," and "EDITOR," plus a four-icon grid of clapperboard mascots captioned "Researcher," "Scriptwriter," "Voice artist," and "Editor." Frame 006 cuts to a yellow "IS IT FACT" title. From frame 007 onward a laptop sits in frame on a teal mechanical keyboard and the overlay switches to the show's scorecard, "☐ FACT OR ☐ HYPE 💩." Frame 007 is the GitHub page for calesthio/OpenMontage, showing "Issues 97," "Pull requests 258," "main," "29 Branches," "0 Tags," commit 9327439 labelled "Merge pull request #527 from rbrogan-git/fix/wan22-t2v-tempora…" dated "3 days ago," "456 Commits," "Watch 350," "Fork 8.2k," an About panel reading "World's first open-source, … video production system. 1… production pipelines, 100+ … 700+ agent skill and produ… knowledge files. Turn your … assistant into a full video p… studio," the link www.openmontage.video, an "AGPL-3.0 license" row, and the counts "64.5k stars," "350 watching," "8.2k forks." Frames 008 through 010 are a Claude Code desktop session titled "OpenMontage installation" reporting "OpenMontage is installed in /Users/tunde/Downloads/open-montage-video and works," "Cloned the repo and ran make setup," "installed the video-rendering parts (Remotion and HyperFrames), installed the offline voice tool (Piper), and created .env from the example file," "Current state: 40 of 137 tools are available, and 1 only partly works," and a cost briefing quoting "'The Last Banana' (60 seconds, 6 Kling clips): $1.33 in total," "Most of the other showcase films: about $4-5 each," and "Cheapest way to get real AI video: one FAL_KEY from fal.ai unlocks several video models." He types "help me set up the fal key." Frame 011 is fal's homepage with the headline "Generative media platform for developers." Frame 012 is fal's Invoices page showing one line, "October 6th, 2026 | paid | $10.00," with a receipt timestamped "October 6th, 2026 at 8:24:20 AM GMT+2." Frames 013 through 015 show the Claude Code home screen reading "What's up next, Tunde?" with lifetime stats "Sessions 470," "Messages 35,582," "Total tokens 2.6B," "Active days 50," "Peak hour 4 PM," "Favourite model Opus 5.5," and the 41-word prompt "Using open montage make a cinematic trailer of a boy named Tunde who sets out to discover which AI tools are fact or hype, losing his friends and everything he owns along the way but never giving up, moody film grain." Frame 016 shows the agent running "Read the mandatory agent guide" and "Read cinematic pipeline manifest and list skills: Failed." Frame 017 shows the delivered cut, "FACT OR HYPE v2 (28s)," with the agent writing "It's 28.3 seconds long, saved as final.mp4" and "0-2.4s: a creator in a ring-lit studio." Frames 018 and 019 are the /usage panel: "Session limit, Resets in 1 hr 37 min, 85%," "Weekly - all models, Resets in 18 hr 47 min, 63%," "Cost $12.53," "API 22m 6s," "Active 45m 31s," "Web searches 1," "Opus 99%," "Haiku 1%," "Cache hit 99%," and a breakdown under "Opus 5.5" of "Input 294," "Output 122.1k," "Cache read 34.1M," "Cache write 309.8k," with the account chip in the footer reading "Tunde Pro." Frames 020 and 021 are fal's usage dashboard: "Total Cost $9.41," "Daily Burn $1.72," "Model API Usage $9.41," "6 endpoints," "Model APIs 100%," itemised as bytedance/seedance-2.0/fast/text-to-video $4.88, fal-ai/kling-video/v3/standard/image-to-video $2.52, fal-ai/nano-banana-pro/edit $1.35, fal-ai/elevenlabs/music $0.60, fal-ai/flux-pro/v1.1 $0.04, and fal-ai/elevenlabs/tts/eleven-v3 $0.02. Frames 022 through 037 play the finished trailer: synthetic creators shouting pieces of "AI just killed video editor," a split-screen of three more, a dim desk scene captioned "EVERY TOOL PROMISED THE FUTURE," a wet night street with a red "HYPE" stamp, an empty room, a red-string conspiracy wall, and the end card "FACT or HYPE."

The repository on screen is real, and its public counters match the GitHub API within display rounding

Frame 007 is not a mockup. Querying the GitHub API for calesthio/OpenMontage returns 64,759 stargazers against the on-screen "64.5k stars," 8,212 forks against "8.2k forks," 353 subscribers against "350 watching," an AGPL-3.0 SPDX license, a homepage of https://www.openmontage.video/, and a default branch of main. The repository was created on 2026-03-29 and last pushed on 2026-10-03.

The commit hash in the frame is checkable too. The head commit is 9327439, dated 2026-10-03T16:28:55Z, with the message "Merge pull request #527 from rbrogan-git/fix/wan22-t2v-temporal-latent." Pull request 527 is titled "fix(comfyui): use a temporal latent in the bundled WAN 2.2 t2v workflow," authored by rbrogan-git, merged at that same timestamp. The screen recording and the repository agree.

The About blurb is truncated in the frame. The full description reads "World's first open-source, agentic video production system. 12 production pipelines, 100+ tools, 700+ agent skill and production-knowledge files. Turn your AI coding assistant into a full video production studio." The phrase the frame cuts off is "agentic," not "AI," and the sentence ends at "AI coding assistant," not at Claude.

Anthropic's own documentation says the "$12 on Claude" figure is an estimate that a Pro subscriber is not billed

The transcript's cost claim is: "So on Claude, this session costs $12 and the video generation costs $9.41." The frame-018 panel does read Cost $12.53. The question is what that number is.

Anthropic's cost documentation for Claude Code answers it directly. The Manage costs effectively page carries this note on the /usage command: "The Session block in /usage shows API token usage and is intended for API users. Claude Max and Pro subscribers have usage included in their subscription, so the session cost figure isn't relevant for billing purposes." The same page adds, "Claude Code computes the dollar figure locally from token counts at list price," and "The figure is an estimate."

The frame establishes that he was on a subscription. The footer chip in frames 013 through 019 reads "Tunde Pro," and the panel above the cost line shows plan meters that only subscribers see: "Session limit, Resets in 1 hr 37 min, 85%" and "Weekly - all models, Resets in 18 hr 47 min, 63%." So the $12.53 measured how much of an already-paid monthly allowance the run consumed. His actual outlay for the whole experiment was the $10.00 fal top-up in frame 012, and $9.41 of that was spent.

The breakdown also explains where the estimate came from. Anthropic's models overview lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, with prompt cache reads priced at 5% of the base input price on that model. The panel reports 34.1M cache-read tokens for the session. At $0.20 per million, cache reads alone account for about $6.82 of the $12.53, which is more than half the estimate. The session ran 22 minutes 6 seconds of API time across 45 minutes 31 seconds of wall-clock activity with a 99% cache hit rate.

Every one of the six fal line items reconciles against fal's published per-second and per-image prices

The six endpoints in frame 020 sum to exactly $9.41 ($4.88 + $2.52 + $1.35 + $0.60 + $0.04 + $0.02). Each line also divides cleanly into fal's listed rate.

fal-ai/kling-video/v3/standard/image-to-video states, "For every second of video you generated, you will be charged $0.084 (audio off) or $0.126 (audio on)." The $2.52 line divides by $0.084 to exactly 30 seconds of audio-off generation.

bytedance/seedance-2.0/fast/text-to-video lists $0.2419 per second at 720p. The $4.88 line works out to roughly 20 seconds. That one endpoint is 52% of the total bill.

fal-ai/nano-banana-pro/edit states, "Your request will cost $0.15 per image." The $1.35 line is exactly 9 image edits. fal-ai/elevenlabs/music states, "$0.6 per output audio minute. The audio will be rounded up to the closest minute," so the $0.60 line is a single music generation. fal-ai/flux-pro/v1.1 is priced at "$0.04 per megapixel," making the $0.04 line one megapixel-sized image.

Put together, the bill bought about 50 seconds of generated motion for a cut that the agent itself describes in frame 017 as "28.3 seconds long." The company behind the dashboard is Features and Labels, trading as fal, and its homepage headline in frame 011 reads verbatim "Generative media platform for developers." The Whisper transcript renders "fal" as "file" throughout, which is a transcription artifact rather than a claim by the host.

"Turns your Claude code into a production studio" names one of five supported hosts, and "world's first" is the project's own wording

The opening line is "This free AI tool turns your Claude code into literally a full production studio." The project's own README states the scope differently: "Works with Claude Code, Cursor, Copilot, Windsurf, Codex — any AI coding assistant that can read files and run code." The repository root confirms this with five separate instruction files and rule sets: CLAUDE.md, CURSOR.md, COPILOT.md, CODEX.md, and .windsurfrules, alongside .claude, .cursor, .codex, and .agents directories. The GitHub topic list includes claude, cursor, copilot, and openai. Claude Code is the host he used, not the only host supported.

The "world's first" line in the About panel is the project describing itself. The README H1 subtitle softens it to "The first open-source, agentic video production system." Open-source pipelines that turn a prompt into a finished short video predate it. The GitHub API dates harry0703/MoneyPrinterTurbo to 2024-03-11 with 128,916 stars and the description "Generate HD short videos from a topic or keyword with an automated AI workflow," and RayVentura/ShortGPT to 2023-06-27 with 8,005 stars. OpenMontage was created on 2026-03-29. Whether it is "first" depends on reading "agentic video production system" narrowly enough to exclude both, which is a definitional argument rather than a verified fact.

The project's current public site makes a different pitch again. openmontage.video describes "Your creative workspace, powered by AI agents," names Claude Code, ChatGPT Codex, and OpenCode as supported agents, lays out a pipeline of "8 stages + 1 gate" (Brief, Research, Script, Scene Plan, The Gate, Narration, Music, Compose, Render), and states that the product is in private alpha with a beta coming. The AGPL-3.0 command-line system on GitHub and the hosted workspace on the site are not the same thing.

The run cost roughly three times what the project's own README budgets for a prompt like his

The video frames $9.41 as the price of finding out. The README publishes its own expectations, and his run came in well above them. Under "With an image/video provider configured" the README quotes a range of "~~$0.15–$1.50," and under "Full setup" it quotes "~~$1–$3." His prompt sits in the "Full setup" band, and it cost $9.41.

The README also documents guardrails that would have flagged this: "Per-action approval — pause for confirmation above a threshold (default: $0.50)" and "Total budget cap — default $10, fully configurable," under a heading that promises "No surprise bills. The agent tells you what it will cost before it spends." He funded the account with exactly $10.00 and spent $9.41, which lands just under the documented default cap.

The showcase costs quoted in frame 009 are accurate to the README. The README's entry for "THE LAST BANANA" reads "a 60-second Pixar-style animated short … 6 Kling v3-generated motion clips (via fal.ai) … Total cost: $1.33," and the two other showcase films are listed at "about $4" and "about $5." So the "$4-5 each" summary the agent gave him matches its source. His own 28.3-second trailer cost more than any of the three published showcase pieces.

One number in the frames does not come from the README. Frame 008 reports "40 of 137 tools are available," which is the local tool registry's own count on his machine. The README claims "100+ production tools" and does not state 137 anywhere.

Key Takeaways

Verified: The repository is calesthio/OpenMontage, licensed AGPL-3.0, created 2026-03-29, with 64,759 stars, 8,212 forks, and 353 watchers at the GitHub API, matching the rounded counters in frame 007.
Verified: The head commit in frame 007, 9327439, is the merge of pull request #527 by rbrogan-git, dated 2026-10-03T16:28:55Z.
Correction: The "$12 on Claude" was not money he paid. Anthropic's cost documentation states that for Max and Pro subscribers "the session cost figure isn't relevant for billing purposes" and that Claude Code "computes the dollar figure locally from token counts at list price." The frames show a "Pro" account chip and subscription limit meters. His out-of-pocket spend for the entire experiment was the $10.00 fal top-up.
Verified: The $9.41 fal total is real and itemised. Six endpoints sum to exactly $9.41, and each line divides cleanly into fal's published rates, including exactly 30 seconds of Kling v3 standard image-to-video at $0.084 per second and exactly 9 Nano Banana Pro edits at $0.15 per image.
Context: A single Seedance 2.0 Fast text-to-video line, about 20 seconds at $0.2419 per second, is $4.88 of the $9.41. Half the bill came from one model.
Partial correction: The tool does not require Claude Code. The README states it "Works with Claude Code, Cursor, Copilot, Windsurf, Codex," and the repository ships a dedicated instruction file for each.
Partial correction: "This free AI tool" and "it's not free of course" are both partly right. The software itself is AGPL-3.0 and free, and the README documents a zero-key path using Piper for narration, Pexels and Pixabay for stock, and local Remotion rendering. The money goes to optional third-party model APIs.
Correction: "World's first open-source … video production system" is the project's own description, not an independent finding. RayVentura/ShortGPT (created 2023-06-27) and harry0703/MoneyPrinterTurbo (created 2024-03-11, 128,916 stars) are open-source prompt-to-video pipelines that predate OpenMontage by two years or more.
Context: The README budgets "~$1–$3" for a full-setup prompt and documents a default $10 total budget cap with per-action approval above $0.50. His run spent $9.41 for a 28.3-second cut, more than any of the three showcase films the README prices at $1.33, about $4, and about $5.
Context: The project's public site describes a different, unreleased thing from the GitHub repository. openmontage.video pitches a "creative workspace" in private alpha with an "8 stages + 1 gate" pipeline and names Claude Code, ChatGPT Codex, and OpenCode.
Unverified: The claim that the delivered trailer is good. That is the host's aesthetic judgement, and frames 022 through 037 are the only evidence either way.
Context: The Whisper transcript writes "fal" as "file" in every instance. The vendor is fal, operated by Features and Labels, and its homepage headline matches frame 011 word for word.

Resources

github.com/calesthio/OpenMontage. The repository in frame 007; its API record confirms the AGPL-3.0 license, the 2026-03-29 creation date, and the star, fork, and watcher counts.
OpenMontage pull request #527. Confirms the head commit shown on screen, "fix(comfyui): use a temporal latent in the bundled WAN 2.2 t2v workflow," merged 2026-10-03.
openmontage.video. The project's public site, which describes a private-alpha hosted workspace with an "8 stages + 1 gate" pipeline rather than the GitHub command-line system.
Claude Code: Manage costs effectively. States that the /usage session cost is a locally computed list-price estimate and "isn't relevant for billing purposes" for Pro and Max subscribers.
Claude models overview. Lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, with cache reads at 5% of the base input price.
fal.ai. The homepage in frame 011, whose headline "Generative media platform for developers." matches the frame verbatim.
fal-ai/kling-video/v3/standard/image-to-video. Publishes $0.084 per second audio off, which divides the $2.52 line into exactly 30 seconds.
bytedance/seedance-2.0/fast/text-to-video. Publishes $0.2419 per second at 720p, placing the $4.88 line at roughly 20 seconds.
fal-ai/nano-banana-pro/edit. Publishes "$0.15 per image," making the $1.35 line exactly 9 edits.
fal-ai/elevenlabs/music. Publishes "$0.6 per output audio minute," rounded up, matching the single $0.60 line.
fal-ai/flux-pro/v1.1. Publishes "$0.04 per megapixel," matching the $0.04 line.
harry0703/MoneyPrinterTurbo. Created 2024-03-11 with 128,916 stars, an open-source prompt-to-video pipeline predating OpenMontage.
RayVentura/ShortGPT. Created 2023-06-27, an earlier open-source framework for automated short-form video.