AI video model and pricing updates

fal lists 50%-off rates through September 30 with stated post-promotion prices; provider credits and Layer Creative Units remain separate raw billing units.

Cloud credit-per-second rates and API USD-per-30-frame rates were verified as separate billing systems. A first-party launch promotion is stated through September 24; no post-promotion price change is inferred.

Artificial Analysis records T2V Elo 1025/rank 28 and I2V Elo 1030/rank 25. The separate 1105.98 model-detail quality Elo remains a different scope.

Current Vidu documentation confirms T2V, I2V, start/end-frame and reference-to-video with synchronized audio, 1–16s main routes and 3–16s R2V. General V2V/editing remains unverified.

Five same-prompt fal-versus-Replicate comparisons and two Runware examples preserve raw provider timings, prompt/settings and example charges.

Live pricing currently lists 12 credits per second at 720p general generation; docs/pricing still lists 11 credits/$0.055. Both official observations and off-peak routes are preserved.

Six access surfaces are mapped across Vidu, Replicate, Runware, fal, MachGen and Roko, with route-specific pricing, duration bands, resolution and billing semantics.

Google token billing and its approximate 720p effective rate are recorded alongside fal per-generated-second tiers, Layer Creative Units and Onysoft per-generation routes without forcing incomparable units into a single cheapest-provider ranking.

Artificial Analysis' Sep-22 image-to-video-with-audio snapshot records Elo 1099, row rank 8, statistical range 7–9 and 4,861 comparisons.

Current xAI documentation confirms T2V, I2V, R2V, first/last-frame controls, video edit and extension, native audio, up to seven reference images and three preset voices.

xAI direct rates are recorded at $0.08/$0.14/$0.25 per generated second for 480p/720p/1080p, with a separate $0.01 input-image charge; gateway per-generation routes remain separate.

xAI, fal, Roko API, reAPI, EvoLink and Runware routes are mapped with model IDs, capability subsets, billing dimensions and example timing where available.

Artificial Analysis snapshots record H3 Max at rank 3 with Elo 1227 in Text-to-Video With Audio and rank 1 with Elo 1195 in Image-to-Video With Audio. The fal internal preference results remain labeled developer-reported.

Current fal, Pixazo, Runware and Layer pricing observations were captured with workflow, resolution and billing-unit boundaries. The comparable fal/Pixazo/Runware 480P and 768P rates are a tie, not a unique cheapest-provider result.

Current H3 Max access and pricing rows were verified across fal, Pixazo, Runware, Cloudflare Workers AI, and Layer. Provider-specific model IDs and raw billing units remain separate.

Artificial Analysis snapshots record HappyHorse 1.1 around rank 8 in T2V with audio and rank 7–8 in I2V with audio, with the observed mode, date, sample size and confidence interval preserved.

Current Alibaba documentation supports HappyHorse 1.1 text-to-video, image-to-video and reference-to-video workflows with native audio, 3–15 second outputs, 24 fps, 480P/720P/1080P tiers and up to nine R2V images.

Alibaba Singapore list rates are recorded at $0.07/$0.14/$0.18 per generated second for 480P/720P/1080P, while Onysoft lists $0.17/$0.22 for 720P/1080P; both remain provider- and region-specific price states.

Alibaba, Onysoft, Modelflare, fal, Segmind and Runware routes were recorded with provider model IDs, workflow scope, resolution tiers, raw billing units and route-specific caveats.

Seven accepted provider-authored examples now preserve prompts, settings, seeds, raw charges or runtime signals; no independent numeric benchmark is claimed yet.

Pruna list rates are recorded across 480P/768P cost, speed and quality modes with derived 50%-off values through September 24; provider example charges and alternative billing bases remain separate.

Pruna, WaveSpeedAI, EachLabs, Runware, Replicate, Scenario, Wiro, Segmind and NanoGPT routes are mapped with model IDs, mode/resolution scope and raw billing or price visibility.

Artificial Analysis snapshots record Wan 3.0 at #1 T2V without audio, #2 T2V with audio, #1 video editing with audio, #6 I2V with audio, and #2 I2V without audio. These are task- and date-scoped observations, not a universal #1 claim.

Alibaba's current guide documents text-to-video, image-to-video, reference inputs, video editing, extension, native audio, and document or webpage inputs under the wan3.0-video Standard surface.

Singapore Standard pricing was recorded as $0.035/$0.07/$0.14 per billable second at 480P/720P/1080P through September 24 00:00 UTC+8, with $0.05/$0.10/$0.20 list rates preserved separately.

Alibaba Cloud Model Studio, Runware, WaveSpeedAI, reAPI, and PixelDojo routes were recorded with provider model IDs, workflow scope, raw billing units, and provider-specific limits.

Google, fal, Layer, Cloudflare Workers AI, Runware and fresh Onysoft routes are mapped to the canonical Google model ID, with workflow, resolution, alias and billing caveats preserved.

Pruna launched P-Video-2-Pro on September 17 as an H3-based text/image-to-video model with generated audio, 480P/768P output and cost/speed/quality modes.

ShengShu released the Vidu S2 real-time video system with S2-Avatar and S2-Editing modes.

Megaton reports a 73.09 index and rank 9 of 16 for H3 Max Turbo, with separate prompt-adherence, scene-consistency, and physics scores. These results are not H3 Max’s older leaderboard row.

Beeble announced SwitchX 2.0 on September 9. This is a release within the SwitchX family, not a model newly launched on September 23.

A dated MaxVideoAI scorecard records an overall 8.3/10 with prompt, visual, motion, consistency, audio, control, stability and value dimensions; it remains attributed editorial evidence rather than a blind scientific leaderboard.

Google's stable Gemini Omni Flash 1.1 replaces the preview surface with documented text-to-video, image-to-video, reference-to-video, edit and extend workflows, native audio, explicit resolution control and a scheduled September 30 preview shutdown.