Guides
Sume guides: the API, Formats, agents, models, media tools and pricing, each answer dated and sourced.
Copilot 'MCP servers in Copilot' policy: admin checklist for Sume
For Copilot Business and Enterprise, an MCP servers in Copilot policy controls remote servers. What an admin checks before allowing Sume's hosted endpoint.
Crop a 21:9 Seedance clip for TikTok: 1120x630 passes, 630x630 misses
A 720p 21:9 Seedance frame is 1470x630. Cropped to 16:9 it is 1120x630, above TikTok's 960x540 floor. Cropped to a square it is 630x630, ten pixels short.
Elysia on Bun: verify a Sume webhook with parse: "text"
An Elysia route that checks the Sume HMAC signature on the raw string, refuses an empty secret, skips duplicates and answers webhook.test. Tested on Bun 1.4.
Price a mixed image batch first: endpoint pricing in Python
Read cost_usd from each Sume image endpoint and multiply by the count. 200 Grok, 50 Seedream 4.5 and 100 Flux 2 Pro images should total $11.25 before you spend.
Wan 3.0 text-to-video: fal list price vs Sume per second
fal lists Wan 3.0 at $0.05, $0.10 and $0.20 per second (read 2026-10-07). Sume bills 25% on top: $0.0625, $0.125, $0.25. Cost of 10 seconds and a minute.
Fashion lookbook video from flat-lay photos: 12 outfits, 9:16 cost
Animate 12 flat-lay outfit photos into 6-second 9:16 clips on Sume: $7.20 on H3 Max, $9.00 on Wan 3.0 or Omni, $41.64 on Seedance 2.5.
FastMCP 4 client OAuth with Sume: request mcp:read, add write later
Use FastMCP's OAuth helper in Python to sign in to Sume's hosted MCP with mcp:read first, then ask for mcp:write only for the script that spends.
FLUX 3 Image 5456x3072 at 300 dpi: print size vs Sume pixel caps
BFL shows FLUX 3 Image at 5456x3072 (16.8 MP). At 300 dpi that prints 18.19 x 10.24 in. Here is what Sume models give at the same dpi.
Google Ads allows ratios between 16:9 and 9:16: Sume models for 3:4
Google Ads lists 16:9, 1:1 and 9:16 and allows ratios between them. Seedance, Wan 3.0 and MiniMax H3 render 3:4 and 4:3 on Sume; Kling 3 and Omni do not.
Google Shorts ad lengths 6-60 s vs Sume model limits: jobs per ad
Google lists 6-60 s for Video Reach and 10-60 s for Video Action, App and PMax Shorts ads. How many Sume jobs a 30 s or 60 s ad takes on each video model.
GPT Image 2.5 image_size auto: priced as 8,294,400 pixels
Sume's docs say image_size auto on GPT Image 2.5 reserves the upper bound of output tokens. Estimate: $0.2224 at high vs $0.0659 at 1024x1024. Send a size.
Grok Imagine returns one image per call: fan out 1,000 in Python
Grok Imagine on Sume takes n=1, so 1,000 images means 1,000 requests at $0.025 each, $25.00 in all. A short Python fan-out shows how to run them safely.
Hotel room tour from two photos: first and last frame API
Turn a door photo and a window photo into a 6-second room walkthrough with frame_images. Ten rooms cost $6.00 to $8.40 on Sume, Seedance 2.5 $34.70.
Interior design mood board to video: up to 10 reference images
Use up to 10 mood-board images and 3 short clips as references for a Gemini Omni Flash 1.1 room video on Sume. An 8 s clip is $1.00 at 720p, $1.50 at 1080p.
A skills repo still recommends Sora 2 and gpt-image-1: swap guide
marketingskills issue 594 lists skill files that name Sora 2 and gpt-image-1. Which lines to edit and which Sume models fit video and image briefs now.
Media tool prices for 10, 100 and 1,000 files on Sume
Cutouts $0.0225, image upscales $0.20, video upscale $0.009 a second, trims $0.02, transcripts $0.01 a minute: totals for 10, 100 and 1,000 files on Sume.
Meta's video from static assets is GA: prepare six photos first
Advantage+ can now generate video from existing images. Prepare the photos so the clip keeps your product intact, and price a backup run on Sume.
MiniMax H3 and H3 Max for TikTok ads: 5 to 15 seconds, priced by tier
MiniMax H3 takes 5 to 15 seconds in 9:16 at 480p or 768p; H3 Max adds 1080p. What 5, 10 and 15 seconds cost on Sume, and the TikTok pixel check to run first.
mulmo images and mulmo movie break on gpt-image-1: set a model now
mulmocast-cli issue 1605: the gpt-image-1 default breaks mulmo images and movie after Oct 23, 2026. Which files to change, and a Python call to Sume instead.
Agents SDK cache_tools_list: stale tools after you grant Sume Write
With cache_tools_list on, an OpenAI Agents SDK MCP server can keep an old tool list. After a Sume Write grant, call invalidate_tools_cache() to see paid tools.
X-OpenRouter-Idempotency-Key vs Sume webhook dedupe on job_id
OpenRouter's webhook dedupe key is job_id plus status. Sume says dedupe on job_id. A SQLite sample builds a job_id plus event key that skips replays.
Roda: verify a Sume webhook with r.body.read in config.ru
A Roda app in one config.ru that verifies a Sume webhook with OpenSSL HMAC and a length-checked secure compare. Aborts on an empty secret. Tested on Ruby 4.0.
SaaS launch teaser from one screenshot: a 12-second AI clip, priced
Animate an app screenshot into a 12-second teaser with first_frame. H3 Max $1.20, Wan 3.0 $1.50, Kling $1.68 sound off, Seedance 2.5 $6.94 on Sume.
Seedream 4.5 costs 1.54x Seedream 4.0 on Sume; 5.0 Lite sits between
On Sume, Seedream 4.0 bills $0.0325, 5.0 Lite $0.04375 and 4.5 $0.05 per image: $32.50, $43.75 and $50.00 per 1,000. The newest is not the dearest.
Transcribing 10, 100 and 1,000 hours: AssemblyAI, Scribe, Sume
Speech-to-text list prices per hour, read 2026-10-07 (AssemblyAI, ElevenLabs, Speechmatics, Cartesia, Gladia) beside Sume at $0.60, for 10 to 1,000 hours.
Split a 31 or 61 second video into valid Sume jobs: Python by model
A 31 or 61 second video cannot be 30 s plus 1 s on Wan 3.0 or Seedance 2.5, since the last job is under the minimum. Split evenly with this Python script.
Submit 6 ad hooks in parallel in Python with one Idempotency-Key each
A Python script that submits six 3-second Omni hook jobs in parallel, each with its own Idempotency-Key so a retry replays the original job. Cost: $2.25.
How many video scenes fit in one Sume script_run? Ten
script_run allows at most 32 paid calls per run. At three paid calls per scene (voice, image, clip) that is ten scenes, with two calls to spare.
Platform presets in Python: build a Sume video request, refuse bad
A 28-line Python preset table for TikTok 9:16, TikTok 1:1 and Shorts ads that picks the lowest Seedance tier clearing the pixel floor and refuses bad lengths.
TikTok Buy Direct: a clip that answers before the one-click tap
TikTok announced Buy Direct checkout from the For You feed on Oct 5. Plan an 8-second clip with size, ship and stock lines for 20 SKUs at $26.00.
Tornado: verify a Sume webhook in a RequestHandler
A Tornado 6.5 RequestHandler that checks the Sume webhook HMAC on self.request.body, raises HTTPError 401 on failure and handles the webhook.test event.
Seedance 9:16 pixel sizes: 496x864, 720x1280, 1080x1920 and cost
Seedance 9:16 frames are 496x864, 720x1280 and 1080x1920: 10,044 to 48,600 tokens a second, so 15 s on Seedance 2.5 costs $4.03 to $21.32.
Split a 181-second Reel into valid AI video jobs (even split)
Greedy 30-second jobs leave a 1-second tail no model accepts. An even split gives 7 Seedance 2.5 jobs or 19 Omni jobs. Python planner and prices at 720p.
Pick a video model for a Reel by length: 3, 10, 15 or 30 seconds
The length you need narrows the Sume models fast: 2 s is Wan only, 10 s allows Omni, 30 s allows Seedance 2.5 or Wan. Table of the shortlist.
AI video generator for TikTok: which Sume models make 9:16
Six Sume video models list 9:16 and one never takes an aspect ratio. Table of ratios, seconds and resolutions, read from the catalog on 2026-10-07.
10-second TikTok hook clip: ten Sume models and their price
For a 10 s vertical hook the price runs from $0.13 on Grok to $5.78 on Seedance 2.5 at 720p. Table of ten models with limits and caveats.
Re-roll cost by AI video job size: 10 s Omni vs 26 s Wan
One bad take costs a whole job again. At 720p that is $1.25 for a 10 s Omni clip, $3.25 for a 26 s Wan clip and $15.02 for a 26 s Seedance 2.5 clip.
Sume video jobs in Go: net/http, a struct, and a poll loop
A Go program that submits a Sume video job with net/http, decodes the job into a struct, polls to completed, and prints the URL and cost in 30 lines.
Generate an AI video with curl and jq in one Bash script
One Bash script that submits a Sume video job with curl, polls until it finishes, and prints the file URL and the cost. A 6-second clip is 75 cents.
TikTok 60-minute uploads vs Sume's 30-minute source limit
Sociality.io reports TikTok uploads up to 60 minutes. Sume trims and Timelines stop at 1800 seconds (30 minutes). What fits, and the cut points for the rest.
Reels vs TikTok length limits and how many AI jobs fill each
Reels 3 min, TikTok 10 min recorded and 60 min uploaded (as reported). On Wan 3.0 that is 6, 20 and 120 jobs; only the first two fit Sume's 30-minute ceiling.
Shorts time-limit option: Timeline render cost at 15 to 180 s
YouTube is testing a Shorts time-limit option. Timeline bills per output minute rounded up, so 60 s costs $0.10 and 61 s costs $0.20.
Cost per kept 8-second clip if you keep 1 in 1 to 1 in 4, by model
Billing is per attempt, not per keeper. Price one kept 8 s clip when you discard 0 to 3 takes first, across six Sume video models. Rates read 2026-10-07.
Self-host an open-weights video model or pay per second?
LTX-2.5 and MiniMax H3 are reported to ship open weights. A break-even formula against Sume's Wan 3.0 rate of $0.125 a second at 720p.
Open-weights video news does not change your Sume bill: LTX, H3, Wan
LTX-2.5 and MiniMax H3 are reported open weights, but Sume still bills hosted per-second rows. Which ids to call and what each costs.
Reels picture-in-picture test: check the four corners first
Instagram is reported to be testing picture-in-picture. Crop each corner of your 1080x1920 Reel with video-filter, free to check, $0.02 to encode.
Picture-in-picture Reel on Sume: what Timeline compose can do
Compose puts one still over one video for $0.02, up to 300 s. Position top, center or bottom; width_ratio 0.05 to 1. It is not a video-in-video PiP.
Go WaitGroup fan-out for Wan 3.0 submits: no mutex, clean -race
Submit 12 Wan 3.0 clips from Go with sync.WaitGroup and a 4-slot channel. Each goroutine owns one slice index, so no mutex. Runs clean under go run -race.