Developers
Building on the Sume API: keys and hosts, jobs and runs, polling, webhooks, idempotency, errors and rate limits, the SDK, the CLI, and MCP auth.
"1080p is not supported by model": which Sume video ids take 1080p
Sume refuses 1080p on rows outside its allowlist, even kling-3 whose catalog lists it. Which ids accept 1080p, which do not, and a client fallback.
MiniMax H3 on Sume: "720p is not supported; use 768p" explained
MiniMax H3 and H3 Max render natively at 480p and 768p. Send resolution 720p and Sume answers 400. Why, what to send, and what it costs per second.
Agent Completion messages[]: system and user turns become one prompt
How Sume joins messages[] into one prompt, what a system turn can and cannot do, and why a GPT-6.1 Sol or Sonnet 5.5 chat history cannot be replayed as is.
Agent Completion output_schema: fail a CI build when it is invalid
Bind output_schema to a Sume Agent Completion and gate CI on the receipt: status completed, output present, output_error empty. Strict schema rules explained.
Cancel an Agent Completion at a deadline: Python poll loop
A Python loop that starts a Sume Agent Completion, polls status_url until next_action stops saying poll_status, and cancels at a deadline.
Chapter markers from Sume STT sentence segments
Request segmentation mode sentence on Sume STT, get time ranges per sentence, and turn the ones you pick into 0:00-style chapter lines with a 12-line formatter.
Opus 5.5 inline tools: adding Sume tools mid-conversation
Claude Opus 5.5 inline tools add tool definitions mid-conversation without breaking the prompt cache. What that changes for Sume MCP, and what it does not.
Sonnet 5.5 on Bedrock has no strict tools: check Sume calls first
Anthropic says strict tool use is unavailable for Claude Sonnet 5.5 on Amazon Bedrock. Validate arguments yourself and use Sume dry_run before paid calls.
Flare draft, Sunburst final: a two-pass image edit loop on Sume
Find the edit on GPT Image 2.5 Flare at low quality, then send the same request to Sunburst at high for the keeper. Costs and limits from Sume docs.
Moving Veo scripts to Gemini Omni: an Interactions API checklist
Veo 3.1 previews shut down Oct 22. Omni uses the Interactions API, task values, base64 or uri delivery and different limits. A checklist for rewriting scripts.
Give every Seedance 2.5 reference a role in the prompt
With up to 12 references per Sume request, say in the prompt what each one is for. Roles for images, videos and audio, and a worked reference-to-video example.
GPT-6.1 Sol function tool for Sume runs: clamp the cap in code
A Responses API function tool that lets GPT-6.1 Sol start a Sume Agent Completion: call_id as Idempotency-Key and a spend cap the model cannot exceed.
GPT Image 2.5 widest banner: 3072x1024 at 3:1, with output cost
GPT Image 2.5 stops at 3:1. 3072x1024 passes the multiples-of-16 and pixel rules; 4:1 and 8:1 banners need another model. Output estimates by quality.
GPT Image 2.5 can take two minutes: handle the 202 job on Sume
OpenAI says complex GPT Image 2.5 prompts can take up to two minutes. Sume blocks 30 seconds, then returns a 202 job. Submit async, poll, and fetch the result.
How big is a 3-minute Short at YouTube's 8 Mbps? About 189 MB
A 180-second 1080p Short at YouTube's 8 Mbps video and 384 kbps audio is about 189 MB. Here is the arithmetic, the TikTok ad caps, and the Sume probe field.
Ideogram 4.5 edits keep the source shape; GPT Image 2.5 needs auto
On Sume an Ideogram 4.5 edit without aspect_ratio keeps the source image's shape, while other edit calls should send aspect_ratio auto. What to send for each.
Instagram 4:5 on GPT Image 2.5: ask for 1088x1360, not 1080x1350
GPT Image 2.5 needs both edges as multiples of 16. 1080x1350 fails that rule; 1088x1360 keeps 4:5 and passes. How to ask on Sume and trim to 1080x1350.
Is GPT Image 2.5 really 50% faster? Time your own calls on Sume
OpenAI calls Flare 50% faster than Sunburst. That compares two models; it is not a promise for your prompt. A small Python timer for your own calls.
Is Kling O3 or LTX-2.5 on Sume? Read the catalog before you plan
New video models appear weekly in vendor feeds. Sume's catalog is one authenticated GET: this checks for Kling O3, LTX-2.5 and Wan 3.0 and reads capabilities.
Kling 3 on Sume rejects reference_*_urls: which model takes references
The kling-3 row supports text-to-video and start/end frames only. Send reference_image_urls and you get a 400; here is where references do work.
One Gemini billing cap pauses every linked project: guard the batch
Google pauses all projects on a billing account at the tier cap. Add a balance check and idempotent submits so one Omni batch cannot stall the rest.
Phone-quality TTS: 8 kHz mu-law on Sume and where MAI Flash fits
Sume's TTS output_format accepts 8000 Hz and pcm_mulaw in wav or raw. What that gives a phone system, what it does not, and how MAI Flash is positioned.
Pick seedance-2 or seedance-2.5 in code by clip length, with price
A small Python helper that sends clips up to 15 s to seedance-2 and longer ones to seedance-2.5, with the 720p price from Sume token math. Runs as written.
Reference image preflight checklist before a Seedance 2.5 job
Most reference-to-video failures are input problems. A checklist and a curl preflight for URLs, count and mix before you pay for a Seedance 2.5 job on Sume.
Replace your mocked Sora client with a Sume contract test
Tests that mocked a Sora client now guard nothing. Write a small pytest contract test around your own video interface, plus one live smoke test against Sume.
Retry a failed image batch on Sume: not billed, same Idempotency-Key
On Sume a failed image generation is not billed, but a retry after a timeout can double-submit. Send an Idempotency-Key per image and retry only the submit.
Write a 30-second Seedance 2.5 prompt as six five-second beats
One 30-second Seedance 2.5 request needs a prompt with a timeline. Split the idea into six beats, keep one subject, and send it as a single Sume job.
Seedance 2.5 camera move test grid: six 4-second probes at 480p
Test dolly, orbit, crane, handheld, whip-pan and static camera prompts with six 4-second Seedance 2.5 probes at 480p for about $6.45 before a 30 s run.
One shared key after Sora? Who can read which Sume job
Sume jobs belong to the member whose key created them. If a worker and a web app use different keys, one gets 404 on the other's job. Plan the key layout.
Turn Sume STT word timings into an SRT file in 30 lines of Python
Sume speech-to-text returns words[] with start and end seconds. Group them into SRT cues of 42 characters or 5 seconds, and see what the caption endpoint takes.
Sume TTS 1.0 rejects model and model_id: use the router to pick
TTS 1.0 has no engine picker and returns 400 for model or model_id. The TTS router takes a required model from its catalog. Compare with ElevenLabs model tiers.
Transcribe a 30-minute file: OpenAI 25 MB cap vs Sume 10-minute jobs
A 30-minute file hits OpenAI's 25 MB upload limit and Sume's 10-minute duration hint. How to split once, transcribe in pieces, and re-base the word timings.
Video filter stops at 300 seconds: which short-video lengths pass?
Sume's video filter refuses sources over 300 seconds. A 3-minute Short passes; a 10-minute TikTok ad source does not and needs a trim first.
AI music longer or shorter than the video: prompt, loop, trim or join
A generated track rarely matches the video length. Four ways on Sume: put the length in the prompt, loop the soundtrack, set an in-point, or join takes.
Captions on a non-English avatar clip: language hint and script_text
For Spanish, French or German scripts, set captions language to a hint or auto and pass script_text so the words match your script, not a mis-heard transcript.
Captions for a clip that switches English and Spanish mid-sentence
MAI-Transcribe-2-Streaming advertises continuous language detection. For a recorded code-switching clip on Sume, lock the wording with script_text.
D-ID API uses Basic auth; Sume uses a bearer API key
D-ID's API docs say requests use the Authorization header with Basic credentials. Sume sends Authorization: Bearer with an API key. A short migration checklist.
Fit a voiceover to a 30-second slot with Sume TTS speed
Measure the first take, divide by the slot length, and set generation_config.speed (0.6 to 1.5). Why a big speed-up is better solved by cutting the script.
Python verifier for Sume webhooks: stale timestamps, empty secret
A short Python function that verifies a Sume webhook: five-minute timestamp window, constant-time HMAC compare and a hard refusal of an empty secret.
Gemini API paid vs unpaid data use: Omni prompts and uploaded clips
Gemini API terms: unpaid content may improve Google products and reach human reviewers; paid content does not. What that means for Omni edit uploads.
HeyGen callback_url vs a signed Sume webhook for a finished video
HeyGen lets you pass callback_url to skip polling. Sume adds mode webhook with a public HTTPS webhook_url and signs each delivery, so verify it.
Image quality defaults differ: OpenAI auto, Sume high, Ideogram medium
Omit quality and the tier differs: OpenAI defaults GPT Image 2.5 to auto, Sume to high, Ideogram 4.5 to medium, Image 1.0 to low. Pin it before you budget.
Lyria 3.5 is single-turn: budget retakes, not edits, on Sume
Google says Lyria 3.5 is single-turn and varies between calls. Plan retakes instead of edits: what 1, 3, 5 and 10 attempts cost on Google and on Sume.
Map your Sora error handling onto Sume's error envelope
Sume errors share one envelope: code, message, request_id. Rebuild your handler around code, rate-limit headers and retry-after, not message text.
MiniMax H3 on Sume: generate_audio false is rejected, audio stays on
The minimax-h3 and minimax-h3-max rows always render native stereo audio. generate_audio false returns 400. How to get a silent clip instead.
No seed on Sume video models: rerun a prompt reproducibly
Sume video models reject a seed field, so a rerun is a new take. For repeatable results store the finished clip, and use Idempotency-Key for safe retries.
OpenAI named no Videos API replacement: build it swappable
OpenAI's deprecation page names no successor to the Sora Videos API. Put video behind one interface, discover models from the catalog, and keep ids out of code.
Mask file checklist for GPT Image 2.5: alpha, same size, under 50 MB
OpenAI says an edit mask must match the image in format and size, stay under 50 MB, and carry an alpha channel. Check your file before sending mask_url to Sume.