Agnes Video 2.5: Cheapest Top-10 AI Video Generator, Pricing & Guide
Agnes Video 2.5 ranks 10th in Artificial Analysis text-to-video (1057 Elo) and costs $2.40 per 1080p minute. Explore its API pricing, benchmarks, and how to start.
👁 118 reads
Regularly updated page. Current as of October 11, 2026.
Agnes Video 2.5 is a video model developed by Singapore-based Sapiens AI. It ranks 10th on the Artificial Analysis text-to-video leaderboard (1057 Elo points, recorded October 11, 2026) and is the only model in the top 10 priced at $2.40 per minute of 1080p video - five times less than the leader, Wan 3.0. A 5-second 1080p clip costs $0.20, while a 10-second clip runs $0.40. In image-to-video, however, performance drops sharply to 25th place.
Current as of October 11, 2026
- Developer: Sapiens AI (Singapore, listed on the App Store as SINGAPORE SAPIENS TECHNOLOGY PTE. LTD.), CEO and founder Bruce Yang.
- AA Ranking: 10th in text-to-video (1057 Elo, 1047-1067 interval), 25th in image-to-video (1031 Elo).
- API Pricing: $0.025/sec in 720P, $0.040/sec in 1080P and 1K, $0.055/sec in 2K.
- Free Tier: Flash version (720P) is temporarily $0, though 503 “video_queue_full” errors have been reported since October 10.
- Where to access: Agnes AI API, Agnes iOS app, Pavo platform, crun.ai aggregator. Not available on fal.ai, Replicate, or OpenRouter.
- Open Weights: Not for 2.5. The older Agnes-Video-2.0-MoE model was released on Hugging Face under the Apache 2.0 license on October 11.
- My own test: Not performed. The API requires an account, which I did not register.
What is Agnes Video 2.5 and who is behind it?
Agnes is a brand owned by Singapore’s Sapiens AI. The company launched its Agnes assistant in 2025, positioning it in press announcements as Singapore’s homegrown answer to DeepSeek. Today, three main products operate under that umbrella:
- “Agnes - Your AI for Life” app on iOS: search, images, video, and presentations.
- API Platform (platform.agnes-ai.com) hosting a family of text, image, and video models.
- Pavo - a short video and micro-drama platform where Agnes Video 2.5 was introduced on August 27, 2026, in a free Flash tier.
The company also builds language models. Agnes 2.5 Pro Beta scored 49 points on the Artificial Analysis Intelligence Index in August, and further 3.0 series models arrived on Hugging Face in October. According to an announcement from March 17, 2026, Sapiens AI raised “tens of millions” of dollars and approached $20M in ARR. Reports vary regarding exact funding totals, so treat those figures with caution.
Video model timeline:
- May 31, 2026 - Agnes-Video-V2.0 announced on X as free.
- July 2, 2026 - First clip from Agnes Video 2.5 Preview shared on X (a horse running through a forest stream).
- August 22, 2026 - Release date of Agnes-Video-2.5 according to Artificial Analysis (no official release statement with this exact date was found).
- August 27, 2026 - Debut on Pavo with the free Flash tier.
- September 25, 2026 - V2.0 deprecated and retired from the API.
- October 11, 2026 - Open weights for Agnes-Video-2.0-MoE released on Hugging Face.
Artificial Analysis leaderboard rankings
Artificial Analysis runs blind side-by-side preference tests: human evaluators view two outputs generated from the identical prompt and pick the superior clip. On September 30, the benchmark updated its text-to-video table to AA-Video-T2V v2.0, where all models are evaluated at 1080p resolution with native audio.
We are launching AA-Video-T2V v2.0, our new benchmark for evaluating text to video models
- Artificial Analysis (@ArtificialAnlys) September 30, 2026
Agnes-Video-2.5 stands 10th with an Elo score of 1057 based on 4,006 pairwise comparisons. Its 95% confidence interval spans 1047-1067, putting it within striking distance of 9th-place SkyReels V4 (1065-1083), but well behind the top eight, which all exceed 1110 points. 11th-place Grok Imagine Video 1.5 sits right behind at 1045 points, meaning Agnes barely holds its top-10 spot.
| T2V Rank | Model | Provider | Elo | USD per minute (AA) |
|---|---|---|---|---|
| 1 | Wan 3.0 | Alibaba | 1156 | 12.00 |
| 2 | Utopai X | Utopai Studios | 1149 | no price |
| 3 | Seedance 2.5 | ByteDance | 1143 | 34.12 |
| 4 | MiniMax H3 (768p) | MiniMax | 1137 | 4.80 |
| 5 | MiniMax H3 Max | MiniMax (via fal) | 1130 | 4.80 |
| 6 | FLUX 3 | Black Forest Labs | 1128 | 17.40 |
| 7 | Seedance 2.0 | ByteDance | 1116 | 22.45 |
| 8 | Gemini Omni Flash 1.1 | 1113 | 9.12 | |
| 9 | SkyReels V4 | Skywork AI | 1074 | 21.00 |
| 10 | Agnes-Video-2.5 | Sapiens AI | 1057 | 2.40 |
| 11 | Grok Imagine Video 1.5 | SpaceXAI | 1045 | 15.00 |
Leaderboard figures from Artificial Analysis text-to-video, recorded October 11, 2026. Pricing reflects the numbers posted on that table.
The chart scale begins at 1000 points, with a full bar at 1160. Under standard Elo formulas, a 99-point gap behind the leader means Agnes wins roughly one out of every three direct head-to-head comparisons against Wan 3.0 (approx. 36%).
Image-to-video (AA-Video-I2V v1.0) paints a much tougher picture: 25th place with 1031 Elo across 6,270 comparisons (1023-1039 interval). The category leader MiniMax H3 Max scores 1195. In that chart, AA quotes the 720p price of $1.50 per minute. If your pipeline relies primarily on starting frames, Agnes is not your best choice.
A quick note on historical figures: Agnes AI’s homepage and some tech outlets mention 1082 Elo, while an older AA table showed 1076. Those numbers stem from previous iterations of the benchmark. Analytics Insight previously reported $0.30 per minute, but that applied to the older V2.0.
Agnes Video 2.5 specifications
| Feature | Text mode | Keyframe mode | Reference mode |
|---|---|---|---|
| Input | prompt only | first and/or last frame + prompt | up to 8 images, 1 video (2-12 s, under 50 MB, 24-60 fps), up to 3 audio tracks (combined 2-12 s); max 12 files |
| Clip length | 4-12 s, default 5 s; one request = one clip | ||
| Resolution | 720P, 1080P, 1K (1024 x 1024), 2K (e.g. 2560 x 1440 in 16:9) | ||
| Aspect ratio | 21:9, 16:9 (default), 4:3, 1:1, 3:4, 9:16 | ||
| Audio | Yes, output includes an audio track; lip sync and spoken languages not specified by developer | ||
| In-prompt references | - | - | <Picture N>, <Audio N>, <Video N> |
| Model identifier | agnes-video-2.5 (Flash: agnes-video-2.5-flash) |
||
Data sourced from Agnes Video 2.5 documentation (checked October 11, 2026). Modes cannot be combined in a single request. Custom frame rates (fps) or frame counts cannot be specified; doing so triggers a 400 error. The Flash tier only outputs 720P, accepts up to 5 images and 3 audio files, and omits reference video support. No hard token limit for prompts is detailed in the docs.
Agnes Video 2.5 pricing breakdown
The rate card is straightforward: billing is strictly per second of generated video, based on resolution. The first 5 input images are included free; additional images cost $0.005 each. Input reference video is billed at the standard per-second rate of your chosen resolution.
| Resolution | USD / sec | USD / min | 5s clip | 10s clip | 12s clip |
|---|---|---|---|---|---|
| 720P | 0.025 | 1.50 | $0.125 | $0.25 | $0.30 |
| 1080P / 1K | 0.040 | 2.40 | $0.20 | $0.40 | $0.48 |
| 2K | 0.055 | 3.30 | $0.275 | $0.55 | $0.66 |
| Flash 720P (promo) | 0 (standard 0.025) | 0 | $0 | $0 | $0 |
Source: Agnes AI pricing page, verified October 11, 2026. Flash is listed as “free for a limited time,” with no announced end date. In practice, compiling a full minute of footage requires at least 5 distinct generations, as single clips max out at 12 seconds.
Alternative platforms and plans
- crun.ai (third-party aggregator): 720P at $0.0156/s, 1080P at $0.0268/s, 2K at $0.0357/s - about 33-38% below official direct rates. A 5-second 1080p generation costs roughly $0.13 there.
- Agnes token plans: Free tier limits users to 1 video request per minute; Starter, Plus, and Pro tiers grant 500 seconds of generation per day. Plan subscription fees are omitted from public documentation.
- Agnes iOS app (US App Store): Starter at $2.99/mo, Plus at $9.90/mo or $99/yr, Pro at $99.90/mo, with credit packs between $0.49 and $2.99. Which underlying model version powers the app is unconfirmed.
- Pavo: Free Flash tier at 720p, 4-12 s, 3 requests per minute, 200 free credits daily (as of August 27, 2026).
For an expanded comparison across alternatives and 100-clip estimates, check Cheapest Top-10 AI Video Generator. For a broader market breakdown, see our guide on how much AI video costs per second.
Cost per minute: Agnes vs. top 10 competitors
| Model | T2V Elo | USD / min | 5s clip | 10s clip | Price multiplier vs. Agnes |
|---|---|---|---|---|---|
| Agnes-Video-2.5 | 1057 | 2.40 | 0.20 | 0.40 | 1x |
| MiniMax H3 (768p) | 1137 | 4.80 | 0.40 | 0.80 | 2x |
| MiniMax H3 Max | 1130 | 4.80 | 0.40 | 0.80 | 2x |
| Gemini Omni Flash 1.1 | 1113 | 9.12 | 0.76 | 1.52 | 3.8x |
| Wan 3.0 | 1156 | 12.00 | 1.00 | 2.00 | 5x |
| FLUX 3 | 1128 | 17.40 | 1.45 | 2.90 | 7.25x |
| SkyReels V4 | 1074 | 21.00 | 1.75 | 3.50 | 8.75x |
| Seedance 2.0 | 1116 | 22.45 | 1.87 | 3.74 | 9.35x |
| Seedance 2.5 | 1143 | 34.12 | 2.84 | 5.69 | 14.2x |
Individual clip costs were linearly scaled from the per-minute figures in the AA table (price / 60 x seconds). Concrete third-party platform rates may differ. The takeaway: Agnes carries the lowest nominal cost, but MiniMax H3 offers an 80-point Elo improvement for double the spend. For many commercial workflows, that jump easily justifies the premium, as analyzed in our head-to-head breakdown.

agnes-video-2.5, POST /v1/videos endpoint, and rates of $0.025 / $0.040 / $0.055 per second.Global availability and access
- Agnes AI API - Sign-up via platform.agnes-ai.com. Documentation notes no regional restrictions for global or EU users. I did not create an account to test sign-up flows firsthand.
- crun.ai - Third-party API reseller with discounted rates.
- Agnes App - Available internationally via the iOS App Store, billed in USD.
- Pavo - Web interface hosted on a .cn domain (app.pavo-ai.cn); I have not tested international accessibility on this service.
- Not available on fal.ai, Replicate, WaveSpeed, or OpenRouter (OpenRouter verified October 11, 2026). As a result, it is not currently included in my Video Arena, which runs through OpenRouter.
Open weights: Closed for 2.5, open for 2.0-MoE
Agnes Video 2.5 remains proprietary and access is strictly via API. However, on October 11, 2026 at 08:50 UTC, Sapiens AI published the weights for an earlier model, Agnes-Video-2.0-MoE, on Hugging Face under an Apache 2.0 license. Key details from the model README (recorded October 11, 2026):
- Built on DaVinci-MagiHuman (GAIR-NLP), featuring 40 layers with SparseMoE blocks (3 experts) on layers 4 through 35;
- 28.7 billion total parameters, with 15.3 billion active during inference;
- Accepts text, video, and audio inputs; outputs video with synchronized audio;
- Weights divided into 7 shards totaling roughly 57.5 GB;
- Requires external dependencies: T5Gemma 9B text encoder (under Gemma license, not Apache), OmniVAE for audio, and the Wan2.2-TI2V-5B VAE;
- Requires Linux and CUDA-capable NVIDIA GPUs; Windows requires WSL.
The repository does not specify minimum VRAM requirements. Because raw checkpoint weights alone take up more than double the 24 GB capacity of an RTX 4090, running it on a single consumer GPU without aggressive offloading to system RAM is impossible. I have not run it locally. For full coverage of the open-source release, read Agnes-Video-2.0-MoE open weights. For self-hosted alternatives, browse our local AI guide.
How to get started with Agnes Video 2.5 step by step
- Create an account on the global developer portal at platform.agnes-ai.com.
- Navigate to the “API Key” section inside the dashboard and select “Create”. Store your secret key in an environment variable rather than hardcoding it.
- Start by experimenting with 720P at 5 seconds ($0.125 per generation) or test the free Flash tier if the queue is open.
- Submit a
POST https://apihub.agnes-ai.com/v1/videosrequest with anAuthorization: Bearerheader. - Extract the
video_idfrom the JSON response and poll the polling endpoint every 1-2 seconds until the status switches tocompleted. Theurlfield returns the final MP4.
Basic curl example from the official documentation (replace the placeholder with your own key):
curl -sS -X POST "https://apihub.agnes-ai.com/v1/videos" \
-H "Authorization: Bearer $AGNES_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "agnes-video-2.5",
"prompt": "At night, three cats march through a forest while playing tiny brass instruments. The camera tracks backward smoothly, moonlight passes through the leaves, with natural footsteps and instrument sounds.",
"seconds": "5",
"mode": "text",
"size": "720P",
"aspect_ratio": "16:9",
"seed": 1101
}'
# Retrieve output
curl -sS "https://apihub.agnes-ai.com/agnesapi?video_id=VIDEO_ID&model_name=agnes-video-2.5" \
-H "Authorization: Bearer $AGNES_API_KEY"
Watch out for field types: seconds must be passed as a string (“5”), not an integer. In keyframe mode, omit aspect_ratio and supply first_frame and/or last_frame image URLs instead. For reference mode, pass arrays under images, audios, and videos. HTTP 429 indicates rate limits, while HTTP 400 points to invalid payload fields.
Prompting best practices
The documentation recommends structuring your prompt using this sequence:
- Subject and setting (who, what, where, when);
- Core action and progression across the scene;
- Camera work (dolly, pan, tracking, framing, focal length);
- Visual styling (lighting, palette, textures, mood);
- Audio design and pacing;
- Consistency constraints (elements that must remain stable).
For official sample prompts across all supported modes, check out What Agnes Video 2.5 Can Do: 8 Real-World Prompts.
Who is Agnes Video 2.5 for?
- High-volume prototyping and rough cuts - 100 5-second generations cost $20 in 1080p and $12.50 in 720p.
- Vertical social clips and micro-content - Native 9:16 support, integrated audio generation, up to 12-second single shots.
- Prompt drafting before premium runs - Test blocking and camera paths cheaply, then send refined prompts to Wan 3.0 or MiniMax.
- Not built for image-to-video pipelines - Ranks 25th in I2V benchmarks; MiniMax H3 Max and Vidu Q4 offer drastically higher visual fidelity.
- Not ready for high-end ad production - The 80-to-99 Elo point deficit against frontier models shows clearly in side-by-side evaluations.
- Not suitable for relying on free Flash in production - Persistent 503 queue errors have been reported since October 10.
Looking for broader options? See our comprehensive AI video generator comparison.
Frequently asked questions
How much does a clip cost with Agnes Video 2.5?
At 1080p, generating a 5-second clip costs $0.20 and a 10-second clip costs $0.40 ($0.040/sec). At 720p, clips cost $0.125 and $0.25, while 2K costs $0.275 and $0.55 respectively.
Can you use Agnes Video 2.5 for free?
The Flash tier (limited to 720P) is temporarily set to $0 on the API and is free via Pavo. However, users have encountered recurring 503 “video_queue_full” errors since October 10 (filed as issue #142 on GitHub, unaddressed by the company through October 11).
Is it the cheapest video generator on the leaderboard?
Within the top 10 of the Artificial Analysis text-to-video table, yes: $2.40 per minute. Cheaper alternatives exist lower down the ranks - such as Boreal at $2.69 per minute - but with a significantly lower Elo score of 838.
Does Agnes Video 2.5 support image-to-video?
Yes, through keyframe mode (first/last frame input) and reference mode. However, it currently ranks 25th on the image-to-video leaderboard.
Does Agnes Video 2.5 generate audio?
Yes, outputs include native audio tracks, and AA benchmarks it on its sound-enabled table. Lip-sync capabilities and non-English speech fidelity are not formally documented. I did not evaluate Polish speech performance.
Can I run Agnes Video 2.5 locally?
Not 2.5. However, you can self-host the open-weights Agnes-Video-2.0-MoE (Apache 2.0 license, roughly 57.5 GB weights, requires Linux and NVIDIA hardware).
Is Agnes Video 2.5 available on fal.ai or OpenRouter?
No (as of October 11, 2026). Outside of the official developer API, it is accessible via the crun.ai aggregator.
Articles in this series
- What Agnes Video 2.5 Can Do: 8 Example Prompts Analyzed
- Cheapest Video Model in the Top 10: Is Agnes Video 2.5 Worth It?
- Agnes-Video-2.0-MoE Released as Open Weights Under Apache 2.0
Sources and review date
- Artificial Analysis Text-to-Video and Image-to-Video Leaderboards (recorded Oct 11, 2026)
- Agnes Video 2.5 Documentation, Flash Tier, Pricing, FAQ, Token Plans, and V2.0 Deprecation (Oct 11, 2026)
- Agnes-Video-2.0-MoE on Hugging Face (README and file manifest, Oct 11, 2026)
- GitHub Issue #142 (Oct 10-11, 2026)
- Agnes on App Store, crun.ai (Oct 11, 2026)
- KuCoin News on Pavo Launch (Aug 27, 2026), VentureBeat (Aug 27, 2026)
- PR Newswire on Agnes Launch (2025), GlobeNewswire Funding Coverage (Mar 17, 2026)
- Agnes AI on X, 2.5 Preview (Jul 2, 2026), Artificial Analysis on X (Sep 30, 2026)


