Agnes Video 2.5 is an asynchronous AI video model with text-to-video, keyframe, and multimodal reference workflows. In our three route checks, its standard text and image-reference jobs completed through GlobalGPT’s Agnes provider. The useful distinction is documented capability: standard 2.5 reaches 720P, 960P, or 2K and accepts video references, while Flash is 720P-only and does not accept reference video.
简答
Choose standard 2.5 when 960P/2K or video-reference inputs matter. Choose Flash only when its 720P limits fit the job and you have checked the current promotion.
Official list price for standard: $0.025/sec at 720P, $0.040/sec at 960P, and $0.055/sec at 2K. Flash documentation lists $0/sec during a limited-time promotion. Platform credits and direct API pricing are separate.
Agnes Video 2.5 is Agnes AI’s OpenAI Videos-compatible model, documented under the model ID agnes-video-2.5. It creates asynchronous video tasks from text, first/last keyframes, or reference media. The standard model supports 4-12 second jobs, 720P, 960P, and 2K output tiers, plus image, audio, and video references. Read the earlier Agnes Video 2.0 guide for context on the product family; this review focuses on the current 2.5 documentation and route checks.
ModesText, keyframe, reference
长度4-12 seconds
输出720P / 960P / 2K
ReferencesImage, audio, video
Agnes Video 2.5 Features: Text to Video, Keyframes, and References
文本转视频
Text mode needs a prompt and supports a chosen duration, resolution tier, and aspect ratio. It is the cleanest starting point for a new scene. For broader tool selection, compare this capability with the workflows in our best AI video generator overview.
Keyframe control
Keyframe mode accepts a first frame, last frame, or both. Agnes documents it for start/end composition control, which is different from simply using a reference image. Use it when the opening or ending composition matters more than a loosely inspired scene.
Image, audio, and video reference
Reference mode lets standard 2.5 use images, audio, or video as content, style, rhythm, or motion input. Agnes explicitly describes reference mode as a workflow that can recompose or retime the result, so it should not be treated as a pixel-preservation tool. For a practical image-to-video framing primer, see how to turn images into cinematic videos.
Hands-on Agnes Video 2.5 Quality Review
We ran three single-run, five-second, 720P, 16:9 tasks through the GlobalGPT/Broly gateway on August 28, 2026. This is a route and output check, not a benchmark. Each delivered file was H.264 with AAC audio at 1280 x 704, 24 fps, and a 5.175-second container duration.
T1, standard text-to-video. Three frames sampled at approximately 0.5, 2.5, and 4.5 seconds.
What our checks support
Text-to-video T1
A rain-wet street, orange lantern, blue-hour light, and forward composition held together across sampled frames.
Image reference T2
The mountain/lake/grass direction remained, but the result was visibly re-styled.
Boundary
T1 introduced unrequested Chinese sign text. These runs cannot establish general text, realism, or motion reliability.
T2, standard image-reference. It supports broad visual direction, not a claim of exact source-composition preservation.
The Flash route accepted the documented agnes-video-2.5-flash model ID and completed through the Agnes provider. Its returned MP4 was byte-identical to T1, so this check cannot support any speed or visual-quality comparison between Flash and standard. We also searched exact 2.5/Flash queries across web, X, YouTube, Reddit, and GitHub but did not locate attributable public reactions in that search. That is a search result, not evidence that no discussion exists.
Agnes Video 2.5 Pricing and Access
Agnes AI’s official documentation gives these direct/API list prices for standard 2.5. The formula is output seconds multiplied by the applicable resolution unit price, with separate documented input-video and excess-image terms where applicable. These are not a quote for GlobalGPT credits or any other platform’s checkout total.
路线
决议
官方价格
Standard 2.5
720P
$0.025 / second
Standard 2.5
960P
$0.040 / second
Standard 2.5
2K
$0.055 / second
闪光灯
720P only
$0 / second during the documented limited-time promotion
For web access, GlobalGPT currently routes standard 2.5 at Agnes Video 2.5 and Flash at Agnes Video 2.5 Flash. Check the live interface for platform credit requirements before you generate. For market context around resolution-based pricing, see Seedance 2.0 的成本.
How to Use Agnes Video 2.5
Choose the standard 2.5 route if you need 960P, 2K, or reference video; otherwise start with the route that matches your available access.
For text to video, write a prompt that states subject, action, setting, camera movement, and any audio intent. Set a 4-12 second duration, resolution, and aspect ratio.
For image to video, use reference mode and refer to the supplied image as <Picture 1> in the prompt. Expect interpretation, not exact preservation.
For a fixed opening or ending composition, use keyframe mode with a first frame, last frame, or both.
Poll the asynchronous task until it is 已完成, then use the returned result URL. Keep API credentials on your server, never in browser code.
Use when quality tier or reference-video support matters
闪光灯
720P only
Up to five reference images
No reference-video input
Comparison limit: our Flash output duplicated the standard test file byte-for-byte. Treat this as routing evidence only, not evidence that Flash is faster, lower quality, or visually identical in normal use.
Who Should Use Agnes Video 2.5?
Use standard 2.5 if you need a documented higher-resolution tier, want to use reference video, or are building a workflow around keyframes and mixed media. It is a sensible candidate for short cinematic scenes where broad visual direction is enough. If 1080P is a hard delivery requirement, check the relevant model’s current documentation rather than inferring it from an adjacent tier; our 1080P support guide illustrates why that distinction matters.
Skip it for now if you need a proven benchmark winner, deterministic typography, or precise image-composition preservation. Our one text run added unrequested sign text, and the reference run noticeably re-styled the scene. Flash is only a fit when 720P, no video references, and a maximum of five images are acceptable.
判决
Agnes Video 2.5 is most compelling as a flexible short-form generation API: the standard route documents text, keyframe, and rich reference inputs across three output tiers. Our controlled checks confirm that both text and image-reference routes completed through GlobalGPT, while also showing why editors should inspect generated lettering and reference fidelity. Its best use is a directed creative workflow with review built in, not a promise of exact control.
Agnes Video 2.5 FAQ
What is Agnes Video 2.5?
It is an asynchronous Agnes AI video model for text-to-video, keyframes, and multimodal reference generation. The standard model ID is agnes-video-2.5.
How much does Agnes Video 2.5 cost?
Official standard list pricing is $0.025/sec at 720P, $0.040/sec at 960P, and $0.055/sec at 2K. GlobalGPT credits are a separate platform price, so check the live route before purchase.
Can Agnes Video 2.5 turn an image into video?
Yes. Standard 2.5 supports image-reference generation, and its keyframe mode can use a first frame, last frame, or both. Reference mode may recompose the source rather than preserve it exactly.
What is the difference between Agnes Video 2.5 and Flash?
Standard supports 720P, 960P, and 2K plus video references. Flash is fixed at 720P, accepts at most five reference images, and does not support reference-video input. The Flash promotion and all platform pricing can change.
