Blog
Sume blog: agent-native video generation, Avatar 1.0 orchestration, and production APIs for builders and coding agents.
Audio edition of a weekly newsletter: 52 issues, TTS year cost
A 1,500-word issue is about 9,000 characters. A year of 52 audio editions costs $10.296 on MAI-Voice-2.1 and $22.36 on Sume TTS, per-job rounding included.
What a 13-point Elo gap means: Wan 3.0 vs Seedance 2.5 win chance
Wan 3.0 leads Seedance 2.5 by 13 Elo on the AA text-to-video board. On the standard Elo scale that is a 51.9% win chance per vote. Arithmetic for other gaps.
What Sume Avatar 1.0 does not do: eight limits to check first
No streaming, no interruption, English-only speech in code, 720p, 4 to 60 seconds, one avatar per video. The limits of Sume Avatar 1.0 in one table.
Which Sume errors mean switch the video model, and which mean wait
Eleven documented Sume error codes and job categories sorted into three actions: fix the request, retry the same model later, or fall back to another model id.
Within 30 Elo of Wan 3.0 for less: MiniMax H3 and H3 Max on Sume
MiniMax H3 (1,137) and H3 Max (1,130) sit 19 and 26 Elo behind Wan 3.0 on the AA board. Sume bills $4.50 and $6.00 a minute at 768p; Wan is $7.50 at 720p.
YouTube end screens need 25 seconds: lengthen a 20 second clip on Sume
YouTube's upload help says a video must be 25 seconds or longer to add an end screen. A 25 second Timeline render with a held still costs $0.10.
14-day holiday countdown from one still: Grok Imagine cost
Fourteen 6-second Grok Imagine Video 1.5 clips cost $1.05 on Sume; joining the 84 seconds adds $0.20. It needs a source image and has no audio.
Retry a timed-out Sume Agent Completion without a second run
Send an Idempotency-Key on POST /v1/agent/completions. A retry returns the original receipt with idempotency_hit true; a changed payload returns a 409.
Agent Completions 403 insufficient_scope: old key or service account?
A 403 insufficient_scope on POST /v1/agent/completions has two causes: a key made before the feature, or a service-account key. details.reason tells which.
Avatar face swap or talking video: recorded presenter vs scripted one
Sume has two ways to put an avatar on screen: face swap onto a 4 to 15 second source clip you recorded, or a script-driven talking video. How to choose.
Sume TTS voice then H3 Max lip sync: 40 seconds in 3 slices
H3 Max lip-sync accepts 5 to 14.8 s of audio. Split a 40 second TTS script into three sentence-based slices at $4.00 on 768p, and what to check.
Black Friday video ads: lock the model id 36 days after Veo ends
Black Friday is Nov 27, 2026, 36 days after Veo 3.1 previews end on Oct 22. A $12 plan: 20 Omni drafts at 360p, 4 finals at 1080p, with dates and arithmetic.
Bulk run spend cap: worst-case ceiling for 100 holiday items
generation_spend_cap_usd is per item, so 100 items at a $4 cap can spend up to $400. The platform maximum is $500 per item; 0 is rejected.
Can I show HunyuanVideo clips to EU viewers? Section 5(c) read
Tencent's HunyuanVideo license says do not use or display Outputs outside the Territory, which excludes the EU, UK and South Korea. Clause by clause.
Can you train a model on open video weights' outputs? Four licenses
H3 and HunyuanVideo bar using outputs to improve other models. LTX bars commercial distillation. Wan 2.2 is Apache 2.0. The clauses, quoted from vendor files.
Swap the LLM behind your MCP client: do Sume's gates change?
No. Hosted MCP gates sit in scopes, idempotency_key and max_spend_usd, not in the client model. What each session type can see, whatever model it runs.
Compliance training narration: 30 modules at 8,000 characters
30 training modules of 8,000 characters are 240,000 characters: $5.28 on MAI-Voice-2.1, $3.60 on Flash, $11.40 on Sume TTS as 30 jobs.
Crop a 10-minute 16:9 ad to 9:16: filter stops at 300 s, so split
Sume video filter takes sources of 300 seconds or less; TikTok non-Spark ads run to 10 minutes. Two trims, two crops and a timeline cost $1.08. The steps.
Do you have to credit an open video model? Label, notice, license
H3 requires a visible 'MiniMax H3' label in commercial products. Hunyuan only encourages 'Powered by'. LTX and Wan ask for notices and a license copy. Per file.
Does Sume offer Haiku 5.5, GLM 5.3 Flash or Mistral Large 4?
Checked against Sume's model catalog on 2026-10-08: Haiku 5.5 and GLM 5.3 Flash have enabled rows; Mistral Large 4, Clef and Strands Decider 2B have none.
Does Sume trim output match YouTube's H.264 and AAC recommendations?
YouTube recommends MP4, H.264 High Profile, closed GOP, AAC-LC. Sume trim docs state libx264, yuv420p and AAC. What matches, what is undocumented, how to check.
Free plan accepts 6 paid jobs: submit 7 and read queue_full
Free allows 1 processing job and 5 queued, so 6 are accepted. A Python script submits 7 in parallel and prints which are accepted and which get 429 queue_full.
Gemini Omni Flash 1.1 at 4K vs 1080p for an 8-second product ad
An 8-second Gemini Omni Flash 1.1 clip costs $1.00 at 720p, $1.50 at 1080p and $3.00 at 4K on Sume. Audio is always on and the ratio is 16:9 or 9:16.
GLM 5.3 Flash JSON output vs Sume's strict output_schema rules
GLM-5.3-Flash supports JSON output. Sume's run output_schema is stricter: object root, additionalProperties false, all properties required. A schema that works.
Griffin-Lite inputs vs a Sume talking-video request, side by side
Tavus describes a reference image plus streaming audio and controls; Sume takes an avatar handle, a script and a tier. The two request shapes compared.
What do 10,000 Haiku 5.5 agent turns cost? 8k vs 120k prompts
Worked token math from Anthropic's Haiku 5.5 price page: 10,000 turns of 8k tokens cost $10.50, while 10,000 turns of 120k tokens cost $625.
HunyuanVideo VRAM: 45 GB at 544p, 60 GB at 720p, 80 GB advised
The HunyuanVideo README lists 45 GB for 544x960 and 60 GB for 720x1280 at 129 frames, tested on one 80 GB GPU. FP8 saves about 10 GB. Table and hosted route.
HunyuanVideo 720p: 1,904 s on one GPU, 338 s on eight
README latency for a 1280x720, 129-frame, 50-step clip: 1,904 s on 1 GPU and 338 s on 8. The GPU-seconds each setup burns, and what it means for hosted jobs.
Imagen 4 with a reference image: Sume rejects it, use Nano Banana 2.1
Sume lists Imagen 4 Fast and Ultra as text-to-image only: input_references is 0 to 0. To edit a photo, use Nano Banana 2.1 at $0.10 per 1K image.
Is a 180 second video still a YouTube Short? Trim to 179.5 for margin
YouTube's Shorts page says up to 3 minutes. It says nothing on rounding, so cut to 179.5 s with Sume video trim (duration accepts decimals, $0.02 a job).
Lip sync a Korean voice recording on Sume: which route, what it costs
Avatar Video is English only in code. For a Korean recording, Sume offers audio-driven routes: H3 Max lip-sync and VEED Fabric. Windows, rates and a test plan.
LTX-2.5 quick start downloads 66 GiB: the five files and a disk plan
The LTX-2 README quick start pulls about 66 GiB for five LTX-2.5 files, including a 12B Gemma text encoder. What to budget, and what Sume lists instead.
MP4 is on both lists: YouTube uploads and TikTok non-Spark ad formats
YouTube accepts 16 formats and TikTok non-Spark ads list five. MP4 and MOV are on both. Sume Timeline and trim return MP4, so one file serves both.
Museum audio guide: 40 stops at 600 characters, voice cost
A 40-stop audio guide at 600 characters a stop is 24,000 characters: $0.528 on MAI-Voice-2.1, $0.36 on Flash, $1.20 on Sume as 40 jobs.
Nightly Sume schedule: model choice and the 30-day cap math
A nightly Scheduled run caps generation at $1.00 by default, so 30 nights is at most $30. How the model's token cost sits beside that cap, with the arithmetic.
Omni references on Sume: billed per output second; Google adds input
Google lists $1.50 per 1M input and $17.50 per 1M output tokens for Omni. Sume bills the provider list x 1.25 per output second. Reference limits compared.
On-hold phone messages: 12 prompts of 300 characters, cost
Twelve on-hold prompts of 300 characters are 3,600 characters: $0.079 on MAI-Voice-2.1, $0.054 on Flash, and $0.24 on Sume as 12 jobs.
One Sume webhook endpoint for job and run events: route on event
Job and run webhooks share one HMAC scheme but not one payload. A 27-line TypeScript handler verifies once, routes on event, and answers 204 to the rest.
Output tokens in a 3 to 10 second Omni clip: 17,376 to 57,920
Gemini Omni Flash bills 5,792 output tokens per 720p second. A 3 s clip is 17,376 tokens and a 10 s clip 57,920. Dollars at Google and on Sume, per length.
Can I use my own LoRA on a hosted video API? Sume has no LoRA field
Sume's /v1/videos has no LoRA, seed or provider-option fields. MiniMax's H3 license allows LoRAs on the open weights. What that means for a character or style.
500-image job on a preview model: Hy Image 3.5 availability, plan B
OpenRouter shows 83.39% availability over 3 days for Hy Image 3.5 Preview, one provider. At 500 images that is about 83 failures to retry. Plan B on Sume.
Stdlib Python Sume webhook receiver: http.server, 204 for webhook.test
A 29-line http.server receiver that refuses an empty secret, checks sume-v1 in a 300-second window, takes the two-signature header, and 204s webhook.test.
A Python Sume webhook verifier for the two-signature rotation header
In a 24-hour secret rotation Sume sends two sume-v1 signatures. This 25-line Python verifier accepts either, refuses an empty secret, and prints its own test.
Can an API caller raise a Sume schedule's spend cap? No
A run request can lower generation_spend_cap_usd below the schedule cap but never above it. The six cases, including null and 0, in one table.
Ratio check before a batch: 4:5, 9:16 and 8:1 on three Sume rows
Imagen 4 lists 5 ratios, GPT Image 2.5 lists 9 and Nano Banana 2.1 lists 15. Which of 4:5, 9:16 and 8:1 each Sume row accepts, and the fallback.
Refresh 1,200 SKU video ads before Black Friday: waves by plan
On an empty workspace, 1,200 jobs take 300 submission waves on Free, 67 on Pro and 14 on Scale, using Sume's wave_size_hint of 75% of accepted capacity.
Re-queue only failed items of a Sume Format bulk queue
A completed bulk queue is not all succeeded. Read counts.failed, pick the failed indexes, and resend them under a new Idempotency-Key. Python, offline.
SaaS release note avatar video: 24 seconds, 3 scenes, cost on Sume
A 24 second release-note video with an AI presenter: hook, demo beat, call to action. Script length, scene plan and cost at standard, plus and max.