Measured runs
Time, cost and every parameter of the video and image runs Orazaka made — per model and provider, read from the run records, with the failures and their causes.
🤖 Generated from code by
scripts/generate-docs.mjs— do not hand-edit. Runorazaka docs buildto refresh.
What the video and image models Orazaka routes to really did, measured on 2026-10-17: each figure below is read from the run records of the platform (start and finish of every Studio run, its size and seed, its outcome) and from the engine's own log — none is typed by hand. Every number is for this setup and is measured again when the setup changes (a new model, engine version, provider or machine).
Setup
| Part | What |
|---|---|
| Machine | MacBook Pro, Apple M1 Max, 64 GB unified memory (Metal) |
| Local engine | stable-diffusion.cpp master-647-72e512a-1-ga397e03 (sd-server, Metal) |
| Platform | Orazaka dev (DigitalOcean App Platform, Toronto) — core, relay, broker; the Mac joins through the GPU relay (chisel over TLS) |
| Hosted provider | DigitalOcean serverless inference (https://inference.do-ai.run) |
| Orazaka commit | dbe135b |
Times are command to file: from the start of the Studio run to the finished output (queueing, transfer and storage included). The engine's own time, where it is logged, is given beside it.
Per model
| Model | Where | Runs (ok / all) | Time, command to file (median, min–max) | Engine time | Price | Spent |
|---|---|---|---|---|---|---|
| Wan 2.2 T2V A14B (video) | DigitalOcean serverless inference | 21 / 21 | 131 s (129–203, n=18) | — | $0.60 per video | $12.60 |
| FLUX.2 [klein] 4B (Q8_0 GGUF) (image) | One Mac (Metal), local | 6 / 18 | 62 s (61–66, n=6) | 72.7 s (60.8–75.0, n=20) | none (your own machine) | — |
| Stable Diffusion 1.5 (image) | One Mac (Metal), local | 6 / 6 | 58 s (58–58, n=6) | 56.6 s (56.5–56.6, n=6) | none (your own machine) | — |
| Stable Diffusion 3.5 Large (image) | DigitalOcean serverless inference | 51 / 51 | 8 s (7–24, n=51) | — | $0.08 per image | $4.08 |
| Z-Image Turbo (Q8_0 GGUF) (image) | One Mac (Metal), local | 0 / 5 | — | 157.8 s (156.7–158.9, n=2) | none (your own machine) | — |
| Wan 2.2 TI2V 5B (image to video, H100 80 GB) (video) | A DigitalOcean GPU droplet, by the hour (through the relay) | 28 / 28 | 177 s (168–279, n=28) | — | $3.99 per GPU hour | $5.93 |
| Wan 2.2 TI2V 5B (image to video, RTX 4000 Ada 20 GB) (video) | A DigitalOcean GPU droplet, by the hour (through the relay) | 1 / 1 | 1106 s (1106–1106, n=1) | — | $0.76 per GPU hour | $0.23 |
| Wan 2.2 T2V A14B (720p, H100 80 GB) (video) | A DigitalOcean GPU droplet, by the hour (through the relay) | 8 / 8 | 1116 s (1104–1302, n=8) | — | $3.99 per GPU hour | $10.24 |
| Wan 2.2 A14B, image to video (720p, H100 80 GB) (video) | A DigitalOcean GPU droplet, by the hour (through the relay) | 7 / 7 | 1179 s (1117–1350, n=7) | — | $3.99 per GPU hour | $9.16 |
3 runs of Wan 2.2 T2V A14B succeeded but are left out of its time statistics (worker restart during the run).
Parameters
Wan 2.2 T2V A14B
DigitalOcean serverless inference · licence Apache-2.0
- size: 832×480 (fixed by the provider: 720×1280, 480×832 and 1280×720 were all answered at 832×480)
- frames: 81
- fps: 16
- duration: 5.06 s
- seed: honoured — four calls with the same prompt and seed gave identical frames
- steps / guidance / scheduler: not exposed by the provider's API
FLUX.2 [klein] 4B (Q8_0 GGUF)
One Mac (Metal), local · licence Apache-2.0
- size: 1024×1024
- steps: 4
- guidance (cfg scale): 1
- sampler: Euler
- scheduler: discrete
- flags: --diffusion-fa
- textEncoder: Qwen3-4B Q8_0
- vae: flux2-vae
- seed: honoured
- memory: 8.3 GB (VRAM)
Stable Diffusion 1.5
One Mac (Metal), local · licence CreativeML OpenRAIL-M
- size: 512×512
- steps: 20
- guidance (cfg scale): 7
- sampler: Euler a
- scheduler: discrete
- seed: honoured
Stable Diffusion 3.5 Large
DigitalOcean serverless inference · licence Stability AI Community Licence
- size: 1024×1024
- steps / guidance / scheduler: not exposed by the provider's API
- seed: not honoured (the provider ignores it)
- output: uncompressed PNG, about 3.1 MB
Z-Image Turbo (Q8_0 GGUF)
One Mac (Metal), local · licence Apache-2.0
- size: 1024×1024
- steps: 8
- guidance (cfg scale): 1
- sampler: Euler
- scheduler: discrete
- memory: 10.5 GB (VRAM)
Wan 2.2 TI2V 5B (image to video, H100 80 GB)
A DigitalOcean GPU droplet, by the hour (through the relay) · licence Apache-2.0
- size: 1280×704 (16:9) and 704×1280 (9:16), the model's native sizes — the photo is cut to the frame around its subject, never letterboxed
- frames: 121
- fps: 24
- duration: 5.04 s (10 s = two linked runs, 238 frames)
- steps / guidance: 40 / 5.0
- seed: honoured
- GPU: NVIDIA H100 80 GB (RunPod), model resident, CPU offload off the 80 GB card
- pipeline: diffusers WanImageToVideoPipeline (the checkpoint's expand_timesteps)
Wan 2.2 TI2V 5B (image to video, RTX 4000 Ada 20 GB)
A DigitalOcean GPU droplet, by the hour (through the relay) · licence Apache-2.0
- size: 1280×704 — native, it fits (19.7 GB of VRAM at the peak) with model CPU offload and VAE tiling; no size ladder step was needed
- frames: 121
- fps: 24
- duration: 5.04 s
- steps / guidance: 40 / 5.0
- seed: honoured
- host RAM: 25.5 GB resident at the peak on a 31 GB host (the image engine and Ollama were stopped)
- GPU: NVIDIA RTX 4000 Ada 20 GB (DigitalOcean droplet)
Wan 2.2 T2V A14B (720p, H100 80 GB)
A DigitalOcean GPU droplet, by the hour (through the relay) · licence Apache-2.0
- size: 1280×720 (16:9) and 720×1280 (9:16)
- frames: 81
- fps: 16
- duration: 5.06 s
- steps / guidance: 40 / 4.0 (high-noise expert) and 3.0 (low-noise expert), boundary 0.875
- seed: honoured
- GPU: NVIDIA H100 80 GB (RunPod), two bf16 experts (57 GB) swapped by CPU offload — about 26 s per step
- load: 129 s from the network volume
Wan 2.2 A14B, image to video (720p, H100 80 GB)
A DigitalOcean GPU droplet, by the hour (through the relay) · licence Apache-2.0
- size: 1280×720 (16:9), the starting image cut to the frame around its subject, never letterboxed
- frames: 81
- fps: 16
- duration: 5.06 s
- steps / guidance: 40 / 4.0 (high-noise expert) and 3.0 (low-noise expert)
- seed: honoured
- GPU: NVIDIA H100 80 GB (RunPod), two bf16 experts swapped by CPU offload — about 28 s per step, 19 to 22 minutes a clip
- checkpoint: Wan-AI/Wan2.2-I2V-A14B-Diffusers (the image-to-video checkpoint of the family, fetched to the pod's own disk: 118 GB in 2 minutes)
Video upscaling
The same clips, taken from 720p (or 480p) to 1080p or 4K by each candidate, on one NVIDIA H100 80 GB. The verdict is a person's, after looking at the same frames side by side with a plain Lanczos upscale; the cost is the run's own time at the GPU's hourly rate.
| Model | Licence | Commercial use | Size | Time (5 s clip) | Cost per clip | Verdict |
|---|---|---|---|---|---|---|
| Real-ESRGAN x4plus (per frame, ffmpeg) | BSD-3-Clause | yes | 1280×704 → 1920×1080 | 104.3 s | $0.12 | sharper than Lanczos but gritty, with a painterly texture on fine detail |
| Real-ESRGAN x4plus (per frame, ffmpeg) | BSD-3-Clause | yes | 1280×704 → 3840×2160 | 127 s | $0.14 | fastest, but the 4K keeps the grain: not chosen |
| SeedVR2 3B (video diffusion) | Apache-2.0 | yes | 832×480 → 1872×1080 | 58 s | $0.06 | winner — clean detail without grain, temporally stable; 480p sources (the Orochia feed) become convincing HD |
| SeedVR2 3B (video diffusion) | Apache-2.0 | yes | 1280×720 → 1920×1080 | 107.4 s | $0.12 | winner — the default of the Studio's upscale: 1080p step |
| SeedVR2 3B (video diffusion) | Apache-2.0 | yes | 1280×720 → 3840×2160 | 280.3 s | $0.31 | winner — 4K holds (tiled VAE, 31 GB): upscale: 2160p, for a hero |
| LTX-2 19B (native high resolution) | LTX-2 Community License | conditions (free under $10M annual revenue) | — | not run | — | not run: weights larger than the pod's free disk; licence conditions to read first |
Image-to-video bake-off
The same FLUX.2 [klein] still made at the video's own aspect through the image-generation Studio (1360×768 and 768×1360; Studio runs c3efac13-6c5b-4743-af3b-e856204f3d96 and 12cee392-9a0e-4a18-a496-44bf0d324c8a), the same motion-only prompt and negatives, the same seed. Direct diffusers runs on the pod, not Studio runs. Landscape: a cable car in the Alps at dawn (a dolly-in is asked); 9:16: a woman and a golden retriever walking in the snow (motion and anatomy). One NVIDIA H100 80 GB; the cost is the run's own time at the GPU's hourly rate. The verdict is a person's, after looking at the frame grid (first, quarter, middle, three quarters, last) of both shots.
| Model | Licence | Commercial use | Shot | Size | Time | Cost | VRAM | Verdict |
|---|---|---|---|---|---|---|---|---|
| Wan 2.2 I2V A14B (40 steps, guidance 3.5) | Apache-2.0 | yes | 16:9 landscape | 1280×720 · 81 fr · 16 fps | 1085.8 s | $1.20 | 35.4 GB | clean and stable, but the camera barely obeys (the cabin recedes instead of the dolly-in); 18 minutes |
| Wan 2.2 I2V A14B (40 steps, guidance 3.5) | Apache-2.0 | yes | 9:16 people | 720×1280 · 81 fr · 16 fps | 1084.8 s | $1.20 | 35.4 GB | the best anatomy: a natural gait, the dog's tail, the coat; no warped limb |
| LTX-Video 0.9.8 13B distilled (8 steps) | LTX-Video Open Weights License 0.X | yes (no revenue clause) | 16:9 landscape | 1216×704 · 121 fr · 24 fps | 13.5 s | $0.01 | 38.7 GB | the winner: sharp, the dolly-in is obeyed, 80× faster than Wan |
| LTX-Video 0.9.8 13B distilled (8 steps) | LTX-Video Open Weights License 0.X | yes (no revenue clause) | 9:16 people | 704×1216 · 121 fr · 24 fps | 14.5 s | $0.02 | 38.8 GB | a good preview: natural walk, but a dark boot-sole artefact on two frames |
| CogVideoX1.5 5B I2V (50 steps) | CogVideoX License (commercial use: a free basic licence after registration, under 1M visits a month) | conditions | 16:9 landscape | 1360×768 · 81 fr · 16 fps | 406.4 s | $0.45 | 26.6 GB | clean but almost static: the requested camera move is ignored |
| CogVideoX1.5 5B I2V (50 steps) | CogVideoX License (commercial use: a free basic licence after registration, under 1M visits a month) | conditions | 9:16 people | — | not finished | — | — | not finished: 768×1360 is refused (tensor size mismatch in the pipeline) |
| HunyuanVideo 1.5 720p I2V (50 steps) | Tencent Hunyuan Community License | no for us: excludes the EU, UK and South Korea, outputs included | 16:9 landscape | — | not finished | — | — | not finished: 19 of 50 steps at 90 s a step (75 GB offloaded), about 75 minutes a clip — and the territory exclusion rules it out for a public site |
| Mochi 1 (genmo) | Apache-2.0 | yes | — | — | not finished | — | — | not run: text-to-video only |
| LongCat-Video (Meituan, 13.6B) | MIT | yes | — | — | not finished | — | — | not run: custom code outside diffusers; to try next |
Winners:
- 16:9 landscape and iteration: LTX-Video 0.9.8 13B distilled — 14 s a 5 s clip at 1216×704, 39 GB, the camera move is obeyed, sharp — and commercial use is free of any revenue clause
- 9:16 people and animals: Wan 2.2 I2V A14B at 40 steps — the best anatomy and gait; 18 minutes, so the finish of a hero clip, with LTX as the 14 s preview
Failures, timeouts and limits
| Kind | What happened | Cause |
|---|---|---|
| access | Wan 2.2 A14B refused the first call (403) | the model access key did not list the model; an all-models key fixed it |
| limit | Wan 2.2 A14B cannot make 9:16 | every size asked (720×1280, 480×832, 1280×720, with or without aspect_ratio/resolution) came back 832×480; the model is refused for 9:16 by name rather than cropped |
| limit | Wan 2.2 TI2V 5B on the Mac (480×832, 25 frames) | MPS out of memory while decoding (26 GiB of MPS, 62 GiB held by other allocations on a 64 GB machine) |
| infrastructure | three video runs took 264 s, 439 s and 1 063 s instead of about 130 s | the Mac worker had lost its tunnel and restarted; the jobs waited in the queue (their times are left out of the statistics) |
| access | openai-gpt-image-2 refused (403) | not available for this subscription tier (tried once) |
| configuration | Stable Diffusion 3.5 Large: first Orazaka runs returned a 1×1 transparent image | the provider's base URL lacked /v1 (the client appends /images/generations): 404, then a transparent fallback; fixed with the provider do-inference-images |
| timeout | FLUX.2 [klein] returned a 1×1 placeholder on several runs (server room and notary desk on every attempt; neon arcade twice) | the image client gives up waiting through the relay (about 2 minutes) while the engine needed 60–75 s plus the transfer of a 2.5 MB image; a web deploy also restarts the relay and drops the tunnel |
| process | two FLUX.2 images (server room, notary desk) were produced in about 62 s each, then lost | later retries wrote placeholders over the same file names before they were published (a process mistake, now guarded by checking the size of each download) |
| timeout | Z-Image Turbo at 1024×1024 was dropped | the engine needs about 157 s per image (8 steps at about 18 s each), longer than the client waits |
| limit | DigitalOcean GPUs: the 48 GB sizes (L40S, RTX 6000 Ada) were never in stock; the 20 GB RTX 4000 Ada appeared and disappeared (3 starts, 2 refused for stock) | sold out in tor1, the only region the private broker's VPC reaches |
| infrastructure | ten clips in a row failed in 17 s (HTTP 500) after the first one succeeded | the worker's free-RAM pre-flight counted the first clip's resident pipeline (~22 GB) against the next job (fixed: no pre-flight for a resident pipeline, orazaka-worker-media#22) |
| infrastructure | three clips (up to 9 minutes of GPU each) were lost mid-render | every deploy of dev.orazaka.com restarts the relay: the worker's AMQP stream dropped and the process exited (fixed: the consumer reconnects and skips a redelivered job, orazaka-worker-media#24) |
| infrastructure | several runs failed with 404 on their starting photo | the platform keeps uploads on the core's disk, which a redeploy wipes; the batch runner uploads the photo again (media durable on Bunny is krizaka/orazaka#144) |
| infrastructure | the droplet's watchdog destroyed an idle droplet (1.2 h) while the laptop that drove the batch was asleep | 30 min without a job; the batch runner now resumes a run in flight from its ticket |
| limit | Wan 2.2 A14B at 720p cannot be resident on one 80 GB card | two bf16 experts are 57 GB; with the activations of 1280×720 the card is full, so the experts are offloaded to the host and swapped (~26 s a step, ~18 min a clip) |
| infrastructure | the RunPod pod kept billing about two hours after its last job | its in-pod watchdog destroys the pod through the RunPod API and the pod has no API key of its own; a guard loop on the maintainer's machine ended it (to fix: a restricted key for the pod, or stop from the driver as soon as the queue is empty) |
| configuration | the first 720p run failed: cached snapshot incomplete | the fetch pattern list had transformer/ but not the second expert transformer_2/ (fixed, orazaka-worker-media#27) |
| quality | three Wan 2.2 TI2V 5B attempts of the mint-tea pour were rejected on review | the stream missed the glass, the glass doubled, the teapot drifted — the subject was replaced rather than a bad clip shown |
| quality | Stable Diffusion 3.5 Large candidates were rejected on review (21 of the 48 images are the ones kept or compared) | a smeared glass-like haze over the face of the tea pourer, a mushy hologram; 1024×1024 is the only size the provider makes, so each kept image is cut to 16:9 |
| access | one finished clip could not be downloaded (404) and was run again | the platform keeps outputs on the core's disk, which a redeploy of dev wipes (durable media: krizaka/orazaka#144) |
| quality | the 720p text-to-video clip « City at night, neon in the rain » was rejected (a neon sign read as gibberish letters despite « no lettering »); « Acoustic guitar by the fire » replaced it | Wan 2.2 A14B draws signage on neon scenes; the text gate caught it on the frame grid |
| quality | the image-to-video clip of the snowy Plateau street drifted in colour on two seeds (grey first frame, then a saturated copper wall); its first ten frames were cut | the model re-exposes the starting image within the first 0.6 s; the rest of the clip is consistent |
| limit | LTX-2 (native high resolution) was not run | its checkpoint is 25 GB in fp8 but the repository publishes a 93 GB text encoder, more than the pod's free disk, and its licence is the LTX-2 Community License (free under $10M annual revenue, a paid agreement above) — a licence to read before it is a default |
| limit | Wan 2.2 A14B image-to-video needs its own 106 GB checkpoint | text and image are two checkpoints of the family; the network volume has 35 GB free, so it is fetched to the pod's container disk each session (2 minutes) instead of a recurring volume |
| process | the first Wan 2.2 bake-off run (18 minutes of GPU) finished but was lost | the bake script saved numpy frames as if they were PIL images (output_type); fixed (output_type=pil) and run again |
| limit | HunyuanVideo 1.5 720p image-to-video was stopped after 19 of 50 steps | 90 s a step with model offload (75 GB resident), about 75 minutes a clip; its licence also excludes the EU, UK and South Korea |
What it cost
| Provider | Output | Price | Runs | Spent |
|---|---|---|---|---|
| DigitalOcean serverless inference | Wan 2.2 T2V A14B | $0.60 per video | 21 | $12.60 |
| DigitalOcean serverless inference | Stable Diffusion 3.5 Large | $0.08 per image | 51 | $4.08 |
| A DigitalOcean GPU droplet, by the hour (through the relay) | Wan 2.2 TI2V 5B (image to video, H100 80 GB) | $3.99 per GPU hour | 28 | $5.93 |
| A DigitalOcean GPU droplet, by the hour (through the relay) | Wan 2.2 TI2V 5B (image to video, RTX 4000 Ada 20 GB) | $0.76 per GPU hour | 1 | $0.23 |
| A DigitalOcean GPU droplet, by the hour (through the relay) | Wan 2.2 T2V A14B (720p, H100 80 GB) | $3.99 per GPU hour | 8 | $10.24 |
| A DigitalOcean GPU droplet, by the hour (through the relay) | Wan 2.2 A14B, image to video (720p, H100 80 GB) | $3.99 per GPU hour | 7 | $9.16 |
| Total of the runs above | $42.24 |
GPU droplet sessions (what was billed)
| Session | GPU | Hours | Rate | Billed |
|---|---|---|---|---|
| DigitalOcean droplet 1 (first start: image build, aborted) | RTX 4000 Ada 20 GB | 0.10 | $0.76/h | $0.08 |
| DigitalOcean droplet 2 (first clip, 18 min; failures) | RTX 4000 Ada 20 GB | 1.91 | $0.76/h | $1.45 |
| DigitalOcean droplet 3 (dropped tunnels, watchdog) | RTX 4000 Ada 20 GB | 1.05 | $0.76/h | $0.80 |
| DigitalOcean droplet 4 (idle while the driver slept) | RTX 4000 Ada 20 GB | 1.20 | $0.76/h | $0.91 |
| RunPod pod (image-to-video pairs, 10 s chains, 720p A14B heroes; it ran ~2 h past its last job because the in-pod watchdog had no key to destroy itself — the local guard loop ended it) | H100 80 GB (CA-MTL-1) | 5.01 | $3.99/h | $19.99 |
| RunPod pod (the 5B image-to-video batch of the home page, the 720p heroes, the first upscaler tests; ended by the spend guard once idle) | H100 80 GB (CA-MTL-1) | 2.50 | $3.99/h | $9.98 |
| RunPod pod (six directed image-to-video clips with Wan 2.2 A14B, one retried, the 1080p/4K passes, the Orochia hero; stopped by id when done) | H100 80 GB (CA-MTL-1) | 2.72 | $3.99/h | $10.85 |
| RunPod pod (the image-to-video bake-off: four models, two shots; a first Wan run finished but crashed in the bake script's save step and was run again) | H100 80 GB (CA-MTL-1) | 1.77 | $3.99/h | $7.06 |
| Total billed | $51.12 |
A droplet is billed from its creation to its destruction: the sessions include the setup (image build, weights), the load of the model and the attempts that failed, which the runs above do not. The cost of a run is its own start-to-finish time at the hourly rate.
The local models cost nothing per output; the energy of the Mac was not measured, so none is claimed. Probes and refused calls (a 403, a size test) are not runs and are not in this table.
See also: Supported models for the registry, and the comparison images on the gallery (/gallery?kind=compare).
Cinematic shorts
2 film(s) made by orazaka cinema (ADR-080, last measured 2026-10-20): script or treatment, character references and keyframes (FLUX.2 [klein]), shots, the cut (SeedVR2 3B, speed ramps, LUT grade, titles, synthesized foley), the voice-over and the music (ACE-Step). Each film's shots, on its GPU:
- NOVA ARDENT: DAWN OF CYAN: Wan 2.2 I2V A14B (4-step distill) on one RunPod H100 80 GB (CA-MTL-1, $3.99/h), upscaled by SeedVR2 3B
- The World of Krizaka: Dawn: Wan 2.2 I2V A14B (40 steps), LTX-Video 0.9.8 13B distilled on one RunPod H200 141 GB (secure cloud, $5.29/h), upscaled by SeedVR2 3B to 2160p (batch 21)
| Film | Frame | Length | Shots (reshot) | Runs (failed) | GPU time of its runs | At the GPU rate | Wall time |
|---|---|---|---|---|---|---|---|
| NOVA ARDENT: DAWN OF CYAN | 1080×1920 (9:16) | 40.5 s | 10 (2) | 35 (5) | 119 min | $7.89 | 237 min |
| The World of Krizaka: Dawn | 3840×2160 (16:9) | 62.5 s | 19 (1) | 20 (0) | 329 min | $29.00 | — |
What it cost per minute of finished film
| Measure | Value |
|---|---|
| Finished film | 1.72 min |
| GPU pod sessions billed (setup, reviews and reshoots included) | $58.01 |
| Billed per minute of finished film | $33.81 |
| The runs' own GPU time per minute of film (what the Studio's estimate multiplies) | $21.50 |
| Wall time per minute of film (one pod, runs in sequence) | 138.3 min |
The billed sessions also rendered the Montréal reel and the news explainer (shots, the reel's cut) that are not delivered yet, and the showcase session's own image-to-video jobs drained by the same pod: the billed figure per minute is this batch's, not a film's. A film's own cost is its runs' GPU time at the rate (the table above). « l'aube » was made outside the Studios on a GPU host running the media worker's own generator and upscaler (orazaka cinema jobs / take), 40-step Wan takes beside concurrent SeedVR2 4K upscales; its keyframes, score and narrator were made on a Mac (Metal, $0).
| Pod session | GPU | Hours | Rate | Billed |
|---|---|---|---|---|
| pod 1 (first start: weights and sd-server build, then released while CI ran) | H100 80 GB | 0.30 | $3.99/h | $1.20 |
| pod 2 (superhero, reel and news shots, two cuts; shared with the showcase session's I2V jobs; stopped by an idle guard at ~14:45, destroyed 15:04) | H100 80 GB | 5.22 | $3.99/h | $20.83 |
| pod 3 (resumed for the reel's music and the news shots; stopped by the coordinator for the GPU queue before a run could start: dev's runs were refused by a saturated billing pool) | H100 80 GB | 0.40 | $3.99/h | $1.60 |
| « l'aube »: H200 141 GB, secure cloud, 03:47–09:47 UTC (stopped by the spend guard at 6 h) | H200 141 GB | 6.00 | $5.29/h | $31.74 |
| « l'aube »: the same pod resumed for one reshoot and three 4K upscales, then terminated | H200 141 GB | 0.50 | $5.29/h | $2.65 |