Guides
Sume guides: the API, Formats, agents, models, media tools and pricing, each answer dated and sourced.
Bulk queue finished_at: measure the wall time of a 100-item batch
A Sume bulk queue sets finished_at when every item is terminal, so created_at to finished_at is your wall time. It does not mean every item succeeded.
One dead image URL fails the whole 100-item Sume bulk create
Sume fetches every attachment at create time, so one broken image URL in a bulk body fails the create before any queue exists. Check URLs first.
Captions for a voiceover: reuse TTS word timings or run STT?
If you generated the voice on Sume, ask for timestamps.words and skip a 1-cent-a-minute STT pass. If the audio is someone else's, run STT. Here is the split.
Claude Code MCP_TOOL_TIMEOUT is about 28 hours; Sume waits 55 s
Claude Code lets a tool call run for about 28 hours by default, but Sume's hosted jobs_wait caps one call at 55 seconds. Why you still loop in short slices.
Claude Code saves MCP text over 50,000 characters: size Sume reads
Claude Code writes long MCP text results to a file, and Sume caps a tool answer at 256 KiB. Read job waves in batches of 20 ids so each answer stays small.
Codex startup_timeout_sec 10 and tool_timeout_sec 60 for Sume
Codex defaults to a 10 s startup and a 60 s tool timeout per MCP server. Sume's jobs_wait holds at most 55 s, so the tool default only just fits.
Cost of 100 eight-second product videos by AI model on Sume
One hundred 8-second clips cost $10 on Grok Imagine, $100 on Wan 3.0 720p and $463 on Seedance 2.5 at 720p. Full model table and plan fit.
One finished minute at three takes a shot: Veo 3.1 vs Omni Flash cost
At three takes per shot a finished minute means 180 generated seconds: $9 on Veo Lite 720p, $21.60 on Fast 1080p, $33.75 on Omni 1080p on Sume.
One finished minute of AI video with voice, music and captions: cost
Six 10-second Wan 3.0 720p clips plus voice, music, Timeline and captions come to about $7.98 a finished minute on Sume. Seedance 2.5 is about $35.
Cursor MCP auth block with CLIENT_ID and CLIENT_SECRET: skip for Sume
Cursor's mcp.json supports an auth block with CLIENT_ID and CLIENT_SECRET for servers that need a fixed client. Sume's hosted OAuth does not; use just the URL.
Do the Sume Agent and API video draw from one credit balance?
Yes. Generation, the Agent, Formats and the API spend one credit pool. Where to see the balance and usage, and what happens when it runs out.
Dub a 12-minute video into 4 languages: STT once, TTS four times
Dubbing a 12-minute video into four languages costs about $1.84 on Sume for transcription and speech, if each script is 9,000 characters. Translation is extra.
Eight concurrent jobs_wait calls per principal on Sume hosted MCP
Sume's hosted MCP bounds waits per process and per principal, workspace plus owner. Batch ids into one jobs_wait instead of sending many parallel waits.
Exact 16:9 sizes on GPT Image 2.5: 1280x720, then steps of 256
Only widths that are multiples of 256 give an exact 16:9 box GPT Image 2.5 accepts. 1280x720 is the smallest; 1024x576 and 1920x1080 both fail.
Test a Sume poll loop against a fake local server in Python
Test your Sume job poll loop with a stdlib http.server that answers queued, processing, completed. No network, no credits and no real waits. Runnable code.
Format io is null on older Formats: pick by your own input kind
On Sume's Format list, the io block is null for Formats saved before the field existed. Do not route on it alone; read input_kind and output_kind when present.
Four 16:9 thumbnail options in one call: n=4 cost, and why n=5 fails
Set n to 4 for four 1536x864 options in one GPT Image 2.5 call: about $0.0420 at medium on Sume. n=5 returns a 400, because every model caps at 4.
Free plan 120 writes a minute: a 100-item bulk queue is one request
A Sume bulk create is one POST against the write budget, and polling spends the separate read budget. Here are the per-plan limits and what that means.
Gemini CLI httpUrl or url for Sume: streaming HTTP versus SSE
In Gemini CLI, httpUrl is for streaming HTTP and url is for SSE. Sume's hosted endpoint is a streamable HTTP server, so use httpUrl with the production URL.
Gemini CLI trust: true removes the prompt; what guards Sume spend
With trust: true, Gemini CLI stops confirming a server's tool calls. For Sume the guards left are idempotency_key, wallet admission and your max_spend_usd.
Omni Flash on Sume: cost of 10, 30 and 60 seconds at every resolution
On Sume, Gemini Omni Flash costs $0.375 to $3.75 for 10 seconds, $1.125 to $11.25 for 30, and $2.25 to $22.50 for 60. Full 360p to 4K table.
Google video resolution upcharge: Standard 0%, Fast 20%, Lite 60%
Going from 720p to 1080p adds 0% on Veo 3.1 Standard, 20% on Fast, 60% on Lite and 50% on Omni Flash on Sume. 4K adds 50% to 200%. Read 2026-10-06.
Tokens in a 1280x720 GPT Image 2.5 image: 106 to 3,787
A 1280x720 GPT Image 2.5 output is 106 tokens at low and 3,787 at max. Token count, dollars at $30 per 1M, and images per million tokens.
GPT Image 2.5 at 2560x1440: 4x the pixels of 720p, about 2x the price
Going from 1280x720 to 2560x1440 quadruples the pixels but only about doubles the GPT Image 2.5 output price at low, medium and high. Table inside.
GPT Image 2.5 at 1536x864: high costs 3.8x medium. Is it worth it?
At 1536x864, high costs 3.9x medium on GPT Image 2.5 ($0.0105 vs $0.0404 on Sume). When to spend it, and when not.
1792x1008 on GPT Image 2.5: the exact 16:9 size between 720p and 1440p
1792x1008 is a legal exact 16:9 size for GPT Image 2.5. About $0.0052 at low to $0.1850 at max on Sume, with the full quality table.
720x1280 vertical 9:16 on GPT Image 2.5: same price as 1280x720
A 720x1280 portrait image costs the same as 1280x720 on GPT Image 2.5, because the estimator uses the pixel count and the ratio. Table by quality inside.
GPT Image 2.5 on Sume: no quality means high, auto reserves max
Skipping quality on GPT Image 2.5 means high ($0.0355 at 1280x720). Sending auto reserves max ($0.1420). What that does to a budget.
Chain two AI clips: last frame in, then first_frame out
To continue a clip on Sume, extract its last frame with video_frames, then send it as first_frame on the next job. It fixes pose and position, not sound.
Does Sume hosted MCP match the HTTP API? Where parity stops
Not fully. Hosted MCP covers most generation families, but Image 1.0 and Video 1.0 are REST-only, and schedules have no MCP tool. Use the HTTP API for those.
How long can an AI video be? Shorts 3 min, TikTok ads 10 min
YouTube allows Shorts up to 3 minutes and TikTok non-Spark ads up to 10. A Sume clip is at most 30 s, but Timeline renders up to 1800 s from several clips.
How long a Sume top-up lasts at 10 AI videos a week on Pro
At 10 fifteen-second Wan 3.0 clips a week, usage runs about $81 a month, so a Pro plan draws about $41 a month from top-ups. A $500 top-up lasts about a year.
Is an idempotency_key an approval step on Sume MCP? No, it is dedup
No. On Sume hosted MCP an idempotency_key stops duplicate submits; it never asks a person. A human check must come from your client, OAuth scope or a spend cap.
Aspect ratios per Sume image model: Banana 2 has 14, Pro 10, GPT 8
Nano Banana 2 lists 14 aspect ratios on Sume, Nano Banana Pro 10 and GPT Image 2.5 8, each plus auto. The full lists, and how to read them from the API.
Is there an ElevenLabs or OpenAI model in Sume's TTS Router? No
Sume's TTS Router v1 lists Cartesia Sonic models only. Other vendors would arrive as new catalog rows. How to check the catalog and what the 400 looks like.
Sume job.completed has no error key; job.failed has payload null
A Sume job.completed webhook omits the error key entirely, while job.failed and job.canceled send payload null plus an error. How to branch without a KeyError.
jobs_wait says status unknown, poll_count 0: is the job lost?
No. A jobs_wait with status unknown and poll_count 0 means the server was too busy to read the job. Wait again on the same ids; never resubmit a paid create.
Kill switch for a Sume video batch: what it stops and what not
A stop file checked between Sume submits halts a batch, but jobs already accepted keep running and billing. A tested Python loop and the cancel limits.
Kling 3 aspect ratios on Sume: 16:9, 9:16 and 1:1 only
kling-3 on Sume accepts 16:9, 9:16 and 1:1. For 4:3, 3:4 or 21:9 pick Seedance 2.5, Wan 3.0 or MiniMax H3 instead. A 400 means the value is not on the list.
Kling 3.0 on Sume: kling-3 is v3 Pro, motion control is separate
On Sume, kling-3 is Kling Video v3 Pro for prompt and frame video; Kling 3.0 Motion Control is a different surface billed on the driving video's seconds.
Localize one Sume Format into six languages: one bulk item each
To make one ad in six languages, send six items to one Sume bulk queue with the language in input and one Idempotency-Key scheme. Here is the shape and limits.
Log format.version on every bulk child: which recipe made which video
A Sume bulk queue can span a Format edit. Store format.version from each child receipt so you know which recipe made which video.
Open-weights video wave: which ids Sume lists, and which it does not
Sume's v1 video catalog lists Seedance, Wan 3.0, MiniMax H3, Kling 3 and Omni 1.1. It has no LTX id and no weights or LoRA fields. How to check before you plan.
Is mcp.dev.sume.com a staging URL I can use? No, use mcp.sume.com
mcp.dev.sume.com is for Sume's own development environments. Customer configs should use https://mcp.sume.com/mcp, and OAuth tokens are bound to that resource.
mcp_health, health_service or health_v1: which Sume ping to call
Three read-only Sume checks answer three questions: mcp_health for your session and tools, health_v1 for the /v1 API path, health_service for plain liveness.
MiniMax H3 Max reference-to-video: what the reserve leaves out
Sume bills H3 Max reference-to-video by output seconds at list x 1.25; the provider's reference-token overage is not reserved. What to check before a big batch.
MiniMax H3 audio references: 2 to 15 s each, 15 s combined
On Sume, MiniMax H3 audio references must be 2 to 15 seconds each, 15 seconds combined, at most 3 files, and never the only reference. Here is how to prep them.
Budget for 10 AI videos a week: 15-second Wan 3.0 clips, 3 resolutions
Ten 15-second Wan 3.0 clips a week cost $9.40 at 480p, $18.80 at 720p and $37.50 at 1080p on Sume. Per-clip, per-month and per-plan numbers.