Blog
Sume blog: agent-native video generation, Avatar 1.0 orchestration, and production APIs for builders and coding agents.
Video filter check passed but the encode failed: what to do next
Sume's free video-filter check covers schema, allowlist and source, not the worker. A valid program can still fail on expressions, memory or time. Next steps.
Which Sume Format for a UGC-style ad? Read io and showcase first
Sume ships ready-made Formats for UGC-style ads. Read each one's io profile and showcase with GET /v1/formats/sume/{slug} before you pick, and cap spend.
What to log for a likeness consent trail: five fields per request
Sume avatar and face swap requests carry an avatar_handle, a source URL, a quality, an Idempotency-Key and a job id. Log those five with the consent record.
Source sizes that reach 3840x2160 within Sume upscale's 4x limit
Sume video upscale accepts factors 1.1 to 4. 960x540 needs 4, 1280x720 needs 3, 1920x1080 needs 2; 854x480 tops out at 3416x1920. Same $0.009 a second.
Sume MCP OAuth in six steps: from the first challenge to a call
Hosted Sume MCP OAuth takes six steps: challenge, authorize redirect, consent page, Write toggle, PKCE code exchange, then a bearer call to the MCP endpoint.
Fit a 330-character script into a 20-second slot: edit, speed, extend
330 characters is about 22 seconds at an assumed 15 characters a second. Fixes on Sume: trim the text, set speed, or lengthen the slot. A take is 1.6 cents.
AI song captions: Lyria lyrics are metadata, so burn your own cues
result.lyrics on a Sume music job is model-reported, not measured. For a lyric video, burn captions with your own cues: music $0.125 plus a $0.20 caption job.
Video inspect seek fast vs precise: stills up to one GOP early
Sume video inspect seek: fast snaps each still to the keyframe at or before the time, about 0-5 s early on typical sources. precise decodes the exact instant.
Cheapest 12-second 9:16 clip for a TikTok ad on Sume, by model
TikTok non-Spark ads need at least 540x960 for 9:16. At 720p or 768p a 12-second clip costs $0.90 on MiniMax H3, $1.50 on Wan 3.0 and Omni, $4.54 on Seedance 2.
12 s of Kling 3.0 Motion Control vs plain Wan 3.0 or Seedance 2.5
Kling 3.0 Motion Control is $0.1575 a second on Sume: $1.89 for 12 seconds, against $1.50 for Wan 3.0 at 720p and $6.93 for Seedance 2.5.
15-second clip: Seedance 2.0 vs 2.5 at 720p and 1080p on Sume
Seedance 2.0 tops out at 15 seconds on Sume. A 15-second clip is $5.67 at 720p versus $8.67 on 2.5, and $12.76 versus $21.32 at 1080p.
22-second Wan 3.0 clip: 720p plus video upscale or 1080p?
On Sume a 22-second Wan 3.0 clip costs $2.75 at 720p, $5.50 at 1080p. Video Upscale adds $0.009 a second, so 720p becomes $2.95. A price check only.
30-second talking video on Sume: Avatar tiers vs Fabric vs H3 Lip Sync
Thirty seconds of talking video costs $5.52 on Avatar Video standard, $7.35 plus, $16.50 max, $5.63 on VEED Fabric 720p and $3.00 on H3 Max Lip Sync at 768p.
47-second voiceover: Gemini 3.8 Flash TTS vs Sume TTS
A 47-second voiceover rounds to 50 seconds on Gemini 3.8 Flash TTS: $0.01125 now, double from 2027-01-01. Sume TTS charges $0.0475 per 1,000 characters.
60-second spot on Sume: six Wan clips, voice, music, render cost split
A 60-second spot of six 10-second Wan 3.0 clips, voiceover, music and render costs $7.77 on Sume. Voice is 0.6% of it, so swapping TTS barely moves the total.
8-second 1080p clip: Gemini Omni Flash vs Seedance 2.5 on Sume
An 8-second 1080p clip costs $1.50 on Gemini Omni Flash and $11.37 on Seedance 2.5 on Sume. Omni Flash 4K is $3.00 for the same eight seconds.
9-second MiniMax H3 vs H3 Max on Sume: 768p and 1080p costs
A 9-second clip is $0.68 on MiniMax H3 at 768p, $0.90 on H3 Max at 768p and $1.80 on H3 Max at 1080p. Only H3 Max has a 1080p tier on Sume.
Cheapest 4-second AI clip on Sume: the minimum-length math
A 4-second clip at 480p costs $1.07 on Seedance 2.5, $0.25 on Wan 3.0 and $0.15 on Gemini Omni Flash at 360p. MiniMax H3 starts at 5 seconds.
14 4K posters: FLUX.3 vs Nano Banana 2.1 vs GPT Image 2.5 on Sume
FLUX.3 4K is $0.607 per image on Black Forest Labs. Sume does not list it. 14 posters cost $8.50 there, $2.80 on Nano Banana 2.1 and $3.12 on GPT Image 2.5.
Generate at 1K and upscale, or generate 4K? Sume price check
Nano Banana 2.1 at 1K plus a Sume image upscale is $0.30, against $0.20 for native 4K. GPT Image 2.5 low plus upscale is $0.22 against $0.22 for high 4K.
A 3-minute YouTube Short from 30-second clips: Wan 3.0 vs Seedance 2.5
Six 30-second clips make a 3-minute Short. Wan 3.0 at 720p is $22.50, Seedance 2.5 is $104.00, plus $0.30 for the Timeline render on Sume.
POST /v1/videos or /v1/video-router/generate for a new build?
Sume recommends POST /v1/videos for new integrations. Video Router stays unchanged with the same models and jobs but a different wire. The differences.
Video frames and reference ingest: no fixed price, a hold that settles
Sume bills video frames and reference ingest by Modal container seconds x list x 1.25 plus a fee, never above the hold. What to expect on the ledger and STT.
Reserved, captured, refunded: the ledger for a 6 s Omni 720p job
A 6-second Gemini Omni Flash 720p job on Sume reserves $0.75, then captures it or refunds it. How to read GET /v1/usage for one job and the three row states.
Claude Dashboards and Motion or a Sume schedule for a weekly recap
For a weekly metrics recap, Claude Dashboards plus Motion fits warehouse data and edits by hand; a Sume schedule fits footage and a $1.00 default cap per run.
Korean captions: slam returns 400, so use korean-ad with language ko
Latin caption styles have no Hangul glyphs, so Korean text on slam, punch or tiktok-green returns 400 caption_hangul_text_latin_style. Use korean-ad.
One webhook route for OpenRouter video and Sume job events
Normalize OpenRouter video.generation.* and Sume job.* webhook bodies to one outcome type, and answer 204 to events you do not know. Runnable Bun/Node code.
OpenRouter signs timestamp,body; Sume signs timestamp.body
The two webhook signatures differ by one separator and a header shape. Side-by-side table and a tested Python verifier for both, refusing empty secrets.
1080x2400 phone recording for Pinterest: crop height to 0.9 or 0.8
A 1080x2400 screen recording is 0.45 wide-to-tall; Pinterest's page avoids shapes shorter than 1:2. Crop height 0.9 gives 1080x2160; 0.8 gives 9:16.
24x36 poster from Nano Banana 2.1 at 4K: 3392x5056, about 140 dpi
Google's size table gives Nano Banana 2.1 at 4K and 2:3 as 3392x5056. On a 24x36 inch print that is about 140 dpi. Sume bills the 4K image at $0.20.
4:5 clip in Snapchat's 720x1280 frame: crop or pad 29.7%
A 1080x1350 clip in a 720x1280 Snap frame: cover loses 304 px of width, contain leaves 380 px of bars. Fit modes cover, contain, blur; $0.10 a minute.
Failed AI video jobs: Google bills successes, Sume refunds the hold
Google says Veo videos are charged only when generated. Sume holds the price at submit, captures it on success and refunds it on failure or early cancel.
Claude Motion makes MP4s from code: when a Sume Agent run fits instead
Claude Motion (beta, 2026-10-08) animates text and charts as code. Sume Agent Completions generate footage and return media URLs. Which fits which job.
createSumeClient sends x-api-key only: an extra header gives 401
The Sume API accepts Bearer or x-api-key but rejects both at once with 401. A fetch wrapper for createSumeClient that drops the extra header, tested offline.
Crop a 1080x1920 clip to a 1080x1080 square for a TikTok 1:1 ad
TikTok's 1:1 ad needs 640x640 or more. A 9:16 source crops to a square with video filter height 0.5625 and y 0.21875, or y 0.1 to keep the top. $0.02.
Divide by 1.25: recover the provider list rate from a Sume video price
Sume bills video at provider list x 1.25. Divide a Sume per-second rate by 1.25 to get the list rate: seedance-2 720p at $0.378 becomes $0.3024.
Edit a video with Gemini Omni Flash on Sume: fields allowed or refused
Omni Flash 1.1 edit mode on Sume takes one video_url plus a prompt, defaults to 720p, and refuses aspect_ratio, duration and other media fields.
Format webhook: 10 s timeout, 10 attempts, about 3 h 5 min of retries
A slow Format webhook gets 10 tries. Without jitter, the last starts 11,010 s after the first; with 10 s timeouts the span is about 3 h 5 min. Code included.
Gemini 3.8 Flash TTS free tier vs paid: test five lines, then budget
Google's pricing page lists Gemini 3.8 Flash TTS as free of charge on the free tier. Sume bills each TTS job, with a 1-cent catalog minimum.
generation_limits missing from a Sume submit: send one, then re-read
The docs say the snapshot is included when Sume can compute it. A Python rule for waves when it is absent, full, or open, tested on three inputs.
GPT Image 2.5 quality auto holds the max price: a 12-image batch
Quality auto on gpt-image-2.5 (Flare, Sunburst) reserves max: $0.21072 of output at 1024x1024 before Sume pricing, so twelve images hold $3.16 or more.
Griffin-Lite waitlist: six checks before planning a live avatar launch
Tavus Griffin-Lite is a research preview with no customer date or price on its page. Six checks before you plan around it, and what to ship on Sume meanwhile.
Sume video prices are provider list x 1.25: checked on Omni Flash
Sume bills video at provider list times 1.25. Dividing Omni Flash's $0.125 by 1.25 gives $0.10 per second, which matches Google's page. Full implied-list table.
Does Kling motion control follow my prompt? Only for appearance
On Sume's Kling 3.0 Motion Control the optional prompt (up to 2,000 characters) steers appearance only; motion and framing follow the driving video.
List Sume video models: /v1/videos/models for limits, /v1/catalog
Two discovery routes, two jobs. /v1/videos/models returns durations, resolutions, aspect ratios and audio flags; the public /v1/catalog lists the wider catalog.
Nano Banana 2.1 vs GPT Image 2.5 on Sume: 4.0x at 1K, 0.9x at 4K
Sume lists Nano Banana 2.1 at $0.10, $0.15 and $0.20 and GPT Image 2.5 at $0.02475, $0.055625 and $0.2225. The price ratio flips below 1 at 4K.
OpenRouter's workspace default callback URL vs Sume's callback_url
OpenRouter has a workspace default for video callbacks. Sume's docs describe callback_url or webhook_url on each request. A Python submit wrapper.
Pinterest wants 6 to 15 seconds: three trims from a 47-second demo
Pinterest's page recommends 6 to 15 seconds with a 4-second minimum. Three video trim jobs at $0.02 turn a 47-second demo into cuts of 12, 9 and 15 seconds.