Blog
Sume blog: agent-native video generation, Avatar 1.0 orchestration, and production APIs for builders and coding agents.

Duration dropdown from supported_durations: 27, 29 and 8 choices
Seedance 2.5 offers 27 lengths, Wan 3.0 29, Omni Flash 1.1 8. Map supported_durations to options in TypeScript, with the price of the longest option.

Spelling bee: 200 word-pronunciation clips on Sume TTS, one job each
200 words of about 12 characters each cost $0.1140 on Sume TTS as 200 jobs, $0.000570 each. One 2,400-character job is $0.1140 but gives one file.

Animate a mascot from one image: H3 start frame or Omni reference
A mascot clip of 10 seconds costs $0.75 as an H3 start frame at 768p, or $1.25 as an Omni reference image at 720p on Sume. The two bodies and the difference.

POST /v1/videos status codes: which of 10 are safe to retry
OpenAPI lists 202 plus ten error codes for POST /v1/videos. Which to retry with the same key, which to fix or stop on, a Python classifier and a backoff plan.

Poll every 2 s or 30 s? Reads per 20-minute video job on Free and Pro
Polling a Sume video job every 2 s is 600 reads in 20 minutes; every 30 s is 40. Share of the read budget on Free and Pro, with a Python check.

Cancel a queued Sume video job: refund rules and the 409 after start
A queued Sume job can be canceled before it starts; after that cancel returns 409 job_generation_already_started. What is held, with a 30-second example.

Node: stream a Sume video to disk with pipeline and a short-file check
Download a finished Sume video from /v1/videos/{id}/content in Node without loading it into memory, and fail if the file is shorter than content-length.

Resume polling Sume video jobs after a restart with a sqlite ledger
Write the job id and Idempotency-Key to sqlite before anything else can crash, then resume polling open jobs on start. Python standard library only.

30-clip storyboard agent: which Sume plan accepts all jobs at once?
If an agent submits 30 separate generation jobs, only Startup (48 accepted) and above take them all at once; Pro (24) needs two waves of 18 and 12.

Edit an existing clip with Gemini Omni Flash 1.1: video_url on Sume
Sume exposes Omni video edit mode through the Video Router video_url field: the prompt is the instruction, resolution defaults to 720p, no duration.

OpenRouter video payload on Sume: drop seed, size, provider.options
Sume /v1/videos follows the OpenRouter shape but rejects seed, size and provider.options with a 400. A Node function that strips them and checks the model id.

ffprobe check for Gemini Omni reference videos: 3 files, 3 s each
Before sending reference videos to Sume's gemini-omni-flash-1.1, check with ffprobe that you have one to three files and each is 3.0 s or shorter. Bash script.

Video Router body to /v1/videos: frame_images, input_references
Map Video Router image_url, end_image_url and reference_image_urls to /v1/videos frame_images and input_references, with a Node converter.

Cheapest 12-second AI video at 720p-class on Sume, five models ranked
For a 12-second clip at about 720p, MiniMax H3 at 768p is cheapest on Sume at $0.90, then H3 Max $1.20, Wan 3.0 $1.50, Seedance 2.0 $4.54, Seedance 2.5 $6.93.

List Sume video model limits with curl, jq and column in one table
curl GET /v1/videos/models into jq and column to see each model's duration range, resolutions and audio flag in one table. Documented limits for six models.

Cheapest AI video API per second: eight listed prices, sorted
The lowest per-second prices on fal, Runway, Google, Replicate and Sume, read 2026-10-09, from $0.025 to $0.09, with which ones Sume actually lists.

FLUX.3 not on Sume yet: detect it via /v1/images/models, fall back
Sume does not list FLUX.3. A Python script reads /v1/images/models, uses FLUX.3 if it appears, else Nano Banana 2.1 at 2K for $0.15. BFL 2k is $0.100.

One ledger for FLUX.3 credits and Sume USD cost with Python Decimal
BFL bills FLUX.3 in credits at $0.01 each; Sume reports usage.cost in USD. A Decimal ledger puts both on one scale: $0.100 + $0.041 + $0.15 = $0.291.

Hosted Sume MCP token hygiene: 5 rules for agent credentials
Five Sume credential rules: an OAuth token is not an API key, stays out of CLI config and prompts, never goes to other providers; rotate exposed keys.

Remote MCP with an API key: x-api-key or Bearer on Sume?
Sume's hosted MCP accepts a key as Authorization: Bearer or as x-api-key. Both see the full tool set; writes and paid calls still need an idempotency_key.

GPT Image 2.5 at 1024px: xhigh is 3,122 output tokens, max is 7,024
Sume's docs give GPT Image 2.5 output estimates at 1024x1024: xhigh $0.09366, max $0.21072 at $30 per million tokens, i.e. 3,122 and 7,024 tokens.

GPT Image 2.5 quality auto reserves max: set quality yourself on Sume
On Sume, GPT Image 2.5 quality auto reserves max. At 1024px the max output estimate is $0.2634 with margin versus a $0.02475 low row. Set quality explicitly.

Runway API vs Sume: six video models, price ratio per second
Runway bills credits at $0.01. Six of its API models have a Sume id. Sume's price is 0.75x to 1.93x Runway's per second; the table gives each ratio.

HEVC phone clip for LinkedIn: exact trim re-encodes, keyframe copies
LinkedIn lists H.264 or VP8. Sume's exact trim re-encodes to libx264, while keyframe trim copies the stream, so a phone HEVC source stays HEVC with keyframe.

Timeline audio mp3 adds padding at every edge: keep wav to rejoin
Sume timeline audio outputs wav (default, sample-exact) or mp3, which adds priming padding at each edge. Re-joined mp3 parts leave gaps.

Replicate FLUX 1.1 Pro vs Sume flux-2-pro: a 250-image refresh
Replicate lists FLUX 1.1 Pro at $0.04 per image. Sume's flux-2-pro is $0.0375. For 250 images that is $10.00 against $9.375; FLUX Schnell is $0.003.

Sora API replacement: four async video APIs and their job shapes
OpenAI's Videos API shut down 2026-09-24 with no replacement listed. OpenRouter, Luma, Runway and Sume all use submit-poll-download; here is each shape.

nano-banana-2 still works but runs as 2.1: check the stored job id
Sume retired Nano Banana 2: google/nano-banana-2 runs as Nano Banana 2.1 and the job stores the 2.1 id. A Python check and what you pay.

One square size for TikTok and LinkedIn ads: 640x640 to 1920x1920
TikTok's square minimum is 640x640 and LinkedIn's maximum is 1920x1920. A 1080x1080 render clears both; 600x600 fails TikTok and 2160x2160 fails LinkedIn.

Video inspect transcribe: no duration hint reserves 1 minute
Sume video inspect with transcribe true bills $0.01 per audio minute. With no duration_seconds it reserves 1 minute; the hint maxes at 600 s (10 min = $0.10).

10-minute ad at 8 Mbps is 600 MB: over TikTok and LinkedIn caps
8 Mbps for 600 seconds is 600 MB. That exceeds TikTok's and LinkedIn's 500 MB caps but fits Pinterest's 2 GB. Sume has no bitrate field, so shorten the cut.

Same idempotency key, new body: Sume Format run 409 vs 200 replay
Same Idempotency-Key and body on a Sume Format run gives 200 with idempotency_hit true. A changed body or attachments gives 409. Bulk replays stay 202.

MAI-Voice-2.1 emotion control vs Sume TTS emotion, speed and volume
MAI-Voice-2.1 lists emotion control. Sume TTS takes generation_config: emotion (1-64 chars), speed 0.6-1.5, volume 0.5-2. A three-take test costs 3 cents.

Runway wan3_prime is 40% above wan3; Sume lists one Wan 3.0 row
Runway lists wan3 at 10 credits a second and wan3_prime at 14. Sume lists wan-3.0 at $0.125 per second at 720p. A 10-second clip: $1.00, $1.40 and $1.25.

Agent run log allowlist: which Sume fields are safe to keep
Log request ids, job ids, status and sanitized media metadata from Sume agent runs; never log API keys, signed URLs, raw private media URLs or full transcripts.

Canceled and skipped Sume agent runs send no webhook: what to poll
A Sume run webhook fires only when a run completes or fails. Canceled and skipped runs send nothing, so your scheduler must read status_url for them.

Nano Banana 2.1, 1,000 images: Sume row vs Google's standard and batch
1,000 Nano Banana 2.1 images at 2K: $150 on Sume's row, $50.40 at Google's standard rate, $25.20 at Google's batch rate, per each page read 2026-10-09.

MCP authorization spec: 6 requirements vs what Sume documents
The MCP authorization spec asks for resource metadata, PKCE and a resource parameter. Here is what hosted Sume MCP documents for each, and what stays open.

Does the Kling motion control prompt change the movement? No
On Sume's Kling 3.0 Motion Control route the motion video drives all movement and the prompt only steers appearance. Fields, limits and $0.1575 per second.

Luma Ray3.2 allows 16 keyframes; Sume offers first and last frame
Luma lists up to 16 keyframes per clip on its Ray3.2 API. Sume does not list Luma; its video models take a first and a last frame through frame_images.

YouTube Short at exactly 60 seconds: 'exceeding one minute' and a trim
YouTube blocks claimed Shorts that exceed one minute. By wording 60.0 s does not, but a 59-second trim leaves margin. Sume trim is $0.02.

Agent Completion attachments: 16 full-size images fit the 500 MB cap
An Agent Completion takes up to 30 images, each up to 30 MB, in a set up to 500 MB. At full size only 16 fit; for 30 images the average must be under 16.7 MB.

Sume Format output_error codes: which to retry and which to fix
output_extraction_failed is transient: re-read the run, then retry with a new key. output_schema_unsatisfied means the schema or recipe is wrong. Six codes.

TikTok ad captions are white and fixed; burned-in captions differ
TikTok's non-Spark page says ad captions show in white with a font you cannot change. Sume's burned-in caption styles live inside the video, a different thing.

Griffin-Lite's 0.27 to 0.59 s latency in 25 fps frames vs Sume clips
Tavus reports Griffin-Lite video latency of 0.43 s average (0.27 best, 0.59 worst). At 25 fps that is 7 to 15 frames. Sume renders recorded clips.

Check a 30-second ad in 15 stills: video inspect fps 0.5 costs nothing
Review a 30-second ad as 15 stills with Sume video inspect fps 0.5. Probe and stills are free; a transcript adds $0.01 per audio minute.

Your agent calls generate_video, or Sume's agent does: two meters
With hosted MCP your model client runs the loop and you pay its provider; with an Agent Completion Sume's agent runs it and you set generation_spend_cap_usd.

Cursor mcp.json for Sume: a URL and nothing else
The Cursor entry for hosted Sume MCP is one url field with no key in the file; OAuth sign-in happens in the client and the consent page lives on the MCP host.