Blog
Sume blog: agent-native video generation, Avatar 1.0 orchestration, and production APIs for builders and coding agents.
Gemini Omni Flash 1.1 ranks 8th, 4th and 3rd on three AA boards
Gemini Omni Flash 1.1 is eighth on AA text-to-video, fourth on image-to-video and third on editing. On Sume it is one id at $0.125 a second at 720p.
Go httptest for a Sume poll loop: assert next_poll_after_seconds
A 29-line Go test: httptest answers in_progress with next_poll_after_seconds 30 twice, then completed, and asserts the loop slept 30 seconds twice.
Go net/http: a Sume video submit retry loop with one Idempotency-Key
A 30-line Go function using net/http that retries a Sume video submit on 429 and 503, reads Retry-After, and reuses one Idempotency-Key. Run on a stub.
Google made Omni Flash its default video model: pin it or sume/auto?
Google's docs now say to use Gemini Omni Flash as the default video model. On Sume you can pin gemini-omni-flash-1.1 or send sume/auto. What each one fixes.
Google says Imagen is shut down: test your Imagen call on Sume now
Google's Imagen page says Imagen models are shut down. Sume lists Imagen 4 Fast and Ultra separately. One catalog check and one request tell you if yours works.
GPT Image 2.5 for packaging text: default high, xhigh and max prices
GPT Image 2.5 on Sume defaults to quality high. At 1024x1024 the provider output is $0.09366 at xhigh and $0.21072 at max, about $0.117 and $0.263 billed.
Griffin-Lite 10 ms audio packets vs Sume's 4 second minimum clip
Two ends of avatar timing: Tavus Griffin-Lite streams audio in packets as small as 10 ms; Sume Avatar 1.0 plans clips of 4 to 12 seconds. What each fits.
H3 on vLLM-Omni: 87 s on 4 B300 for one clip, and the cost
MiniMax says an 8.7 s H3 clip takes about 87 s on 4 B300 GPUs. Turn that into GPU-seconds, a break-even rate, and compare with a 9 s Sume job.
HappyHorse 1.1 is 12th at 1,042 Elo: what ranks above it on Sume
HappyHorse-1.1 is 12th on the AA text-to-video board at $10.80 a minute. Sume does not list it; six Sume models rank above it. Per-minute prices inside.
Hosted video instead of a GPU: a Python submit-poll-download script
Replace a local H3 or Wan pipeline with three HTTP calls. A Python script that submits a minimax-h3 job to Sume, polls it and saves the MP4, with job states.
Is Wan 3.0 an upgrade on Wan 2.7? 113 Elo more on the AA editing board
On the AA video-editing board Wan 3.0 (1,187, $12.00 a minute) beats Wan 2.7 (1,074, $16.90) by 113 Elo. Sume lists Wan 3.0, not 2.7; it runs 2 to 30 s.
Kling 3.0 Omni is 9th for video editing at 1,000 Elo: Kling 3 on Sume
Kling 3.0 Omni (1080p Pro) is ninth on the AA video-editing board at $10.91 a minute. Sume lists kling-3 for generation: $8.40 or $12.60 a minute, 4 to 15 s.
Luma Ray 3.2 HDR costs 2x SDR and EXR 3x: 5 s at 720p and 1080p
Luma lists a 5 s Ray 3.2 clip at $0.30 SDR, $0.60 HDR and $0.90 HDR plus EXR at 720p; $1.20, $2.40 and $3.60 at 1080p. Sume does not list Luma.
Luma Ray 3.2 video-to-video: 20 s, 64 keyframes, vs Omni edit on Sume
Luma's Ray 3.2 video-to-video keeps the source length up to 20 s and takes up to 64 keyframes. Sume does not list Luma; its edit route is Gemini Omni Flash 1.1.
Mac or Windows team: which open video models need Linux and CUDA?
HunyuanVideo and HunyuanVideo-1.5 list Linux; LTX-2.5's fastest decoder is Linux plus CUDA; H3 has a Mac route. What each page says and a hosted option.
MiniMax H3 4K hero clip vs 768p cutdowns: holiday ad cost
A 10-second MiniMax H3 clip is $0.75 at 768p, $1.625 at 2K and $2.00 at 4K on Sume. Five free reference images are included; extra ones are $0.10 each.
MiniMax H3 on a 12 GB RTX 3060: 480p only, 42 GB, and the price
MiniMax says 12 GB of VRAM runs 480p H3 clips with audio from a pruned int8 checkpoint of about 42 GB. What that buys, and a 480p clip on Sume costs.
minimax-h3 vs minimax-h3-max on Sume: price per clip at 768p and 1080p
At 768p, minimax-h3-max costs $0.10 a second on Sume against $0.075 for minimax-h3. 1080p exists only on Max. Per-clip prices for 5, 10 and 15 seconds.
Omni Flash forecast: 200, 500 or 2,000 eight-second clips a month
A monthly forecast for gemini-omni-flash-1.1 on Sume: 200, 500 or 2,000 eight-second clips cost $60 to $6,000 depending on resolution. Four-column table.
Nano Banana 2.1 4K batch: handle 200 and 202, reuse the key
A 4K Nano Banana 2.1 call can return 200 with the image or 202 with a job. A short Python handler, the retry rule, and what 4K costs: $0.20 per image on Sume.
Nano Banana 2.1 4K is 3.4x the 1K price at Google, 2x on Sume
Google bills 1,120 tokens at 1K and 3,780 at 4K for Nano Banana 2.1, a 3.375x gap. Sume's tiers go $0.10 to $0.20, 2x. Why the ratio differs.
A news app's Listen button: 5,000 backlog articles, voice cost
Voicing a backlog of 5,000 articles at 4,000 characters is 20 million characters: $440.00 on MAI-Voice-2.1, $300.00 on Flash, $950.00 on Sume TTS.
Nine Performance Max videos (3 ratios x 3 concepts): Sume cost
Three concepts in 16:9, 1:1 and 9:16 at 10 seconds each is 90 video seconds. At Wan 3.0 720p that is $11.25 on Sume; Seedance 2.5 is about $52.
Omni Flash resolution ladder on Sume: 0.3x, 1x, 1.5x, 3x of 720p
On Sume, Omni Flash 1.1 costs 0.3x at 360p, 1.5x at 1080p and 3x at 4K compared with its 720p rate of $0.125 a second. Per-clip prices at 8 and 10 seconds.
One Sume avatar handle across talking video, TTS, stills and lip sync
A Sume avatar handle is accepted by talking-video, TTS voice selection, face-in-still images and the lip-sync routes. What each uses from it and what it costs.
One idempotency key per model when a video fallback changes payload
Reusing an Idempotency-Key after switching video models returns 409 idempotency_conflict. Build the key from order, model, and prompt hash so retries stay safe.
Packaging label text test: 12 prompts on 3 Sume models, what it costs
Test Nano Banana 2.1, Imagen 4 Ultra and GPT Image 2.5 on 12 label-text prompts first: about $1.20, $0.90 and $1.40 on Sume. A short script runs it.
Python argparse CLI that checks seconds per Sume model first
A 25-line argparse CLI: choices reject unknown models, a duration check uses the documented limits, and --dry-run prints the body. Runs offline.
Python fallback chain for Sume video models: 404 and 503 only
After the Sora API shutdown, a model chain must not retry everything. This Python function moves on at 404 and 503 and stops on 429, 402, 400, 409.
Parse Retry-After as seconds or HTTP-date before retrying a Sume 429
A 21-line Python helper that reads retry-after as integer seconds or an HTTP-date, caps the wait, and falls back to exponential delay when the header is absent.
Python urllib: retry a Sume video submit on 429 and 503, no requests
A standard-library Python function that retries a Sume video submit on 429 and 503, reusing one Idempotency-Key and reading Retry-After. Tested on a stub.
A 2,000-word blog post read aloud: MAI-Voice vs Sume TTS cost
A 2,000-word post is about 12,000 characters: $0.264 on MAI-Voice-2.1, $0.18 on Flash, $0.57 on Sume TTS in one request. Arithmetic shown.
Seedance 2.5 leads 2.0 by 27 Elo on text-to-video, not image-to-video
Seedance 2.5 leads 2.0 by 27 Elo on AA text-to-video and 75 on editing, but 2.0 is fifth on image-to-video and 2.5 is not in the top 12. Sume prices both.
Seedance 2.5 or Wan 3.0 for text-to-video: 13 Elo, 2.8x the price
Wan 3.0 beats Seedance 2.5 by 13 Elo on text-to-video and 36 on editing, at $12.00 against $34.12 a minute. Pick by length, price and inputs.
Shopify image limit 20 MB / 25 MP vs GPT Image custom sizes on Sume
Shopify caps images at 20 MB and 25 MP. Sume's GPT Image custom size tops out at 8,294,400 px (3840 edge), so file size, not megapixels, is what you hit.
Shopify's 5 recommended video ratios vs what each Sume model makes
Shopify recommends 16:9, 9:16, 4:3, 3:4 and 1:1. Seedance, MiniMax H3 and Wan 3.0 cover all five; Kling 3 lacks 4:3 and 3:4, and Gemini Omni Flash only has 2.
Does Sume Avatar 1.0 speak Spanish or Korean? English only in code
The Avatar 1.0 talking-video prompt on main says English only. What that means for Spanish or Korean lines, and the audio-driven route to test instead.
Keep untrusted text out of the prompt: Sume run input is data only
Sume writes a run's input object to a file and treats it as data, not instructions. Schedule runs allow 64 properties and 2 MiB. Use it as a guardrail.
Test an avatar's look with a 4 second clip before a 60 second run
A 4 second Sume clip costs $0.736 on standard; a 60 second one $11.04. Use the cheap run to check face, voice and framing first. Numbers and checklist.
Test Sume retry code with a unittest fake server: 429 then 202
A stdlib fake server answers 429 with Retry-After, then 202. The test asserts your retry reuses the same Idempotency-Key and waits 7 seconds, with no sleep.
AA text-to-video top 12: six you can call on Sume, six you cannot
Of the 12 models on the AA text-to-video board, six have a Sume id that takes a prompt-only request. The list, with Elo and Sume's price per minute.
Black Friday supercut in Timeline: 12 segments vs auto chunking
Timeline accepts up to 200 video slots. A single strategy is refused above 12 segments (render_strategy_unsafe); auto chunks the render past 12.
Mixing 24 fps Seedance clips with 30 fps footage in Sume Timeline
Set output.fps to 24, 25, 30 or 60 in Timeline 1.0. If you omit it, the job matches the sources, and a mismatch returns output_fps_resamples_sources.
A Veo 3.1 call becomes a Sume Omni job in under 30 lines of Python
Replace a Veo 3.1 request with a Sume gemini-omni-flash-1.1 job: submit, poll every 30 seconds, download the mp4 and read usage.cost. Standard library only.
Veo 3.1 price as a multiple of Sume's Omni rate: 3.2x, 0.8x, 0.4x
At 720p Veo 3.1 Standard is 3.2x Sume's Omni rate, Fast is 0.8x and Lite 0.4x. The ratios at 1080p and 4K are in one table, with the Oct 22 caveat.
Video editing board: Wan 3.0 1,187, Seedance 2.5 1,151; Sume uses Omni
Wan 3.0 tops the AA video-editing board at 1,187 Elo. Sume's documented edit route is video_url on gemini-omni-flash-1.1, third at 1,139. Price inside.
Wallet reserve for a full queue of Seedance 2.5 clips, by plan
Sume reserves the estimate at submit. A full queue of 5 s Seedance 2.5 720p clips ($2.889 each) reserves $17.33 on Free and $346.68 on Scale.
Wan 2.2 CUDA out of memory: offload_model, dtype and t5_cpu flags
The Wan 2.2 README fixes OOM with three flags: offload_model, convert_model_dtype and t5_cpu. Which commands need 80 GB and 24 GB, and what it does not promise.