最近Agnes AI,开始出现在开发者和创作者的讨论里。
原因就很直接,这家厂商:把 API 全部免费开放了。
文本、图片、视频,全模态,全免费
是的,没有试用额度,没有阶梯付费,没有什么门槛和废话,就是直接可用。
这件事如果只看表面,会觉得是“又一家做免费策略”的厂商,市面上免费的很多,好用的太少了。
但如果看它最近一周的调用数据,会发现事情不一样。
单周 4.66 万亿 Token
跑出的真实“压测”成绩单
Agnes AI 官方最新公布的数据里,Agnes 全模态模型单周总调用量已经到了 4.66 万亿 Token:
单看4.66 T(万亿)这个数字,你可能没有直观感受。
这个调用量,仅次于 DeepSeek V4 Flash 在 OpenRouter 上的单周调用量,这足以体现它当下的高服务规模。
按 1 个 Token 约等于0.8个汉字粗略换算,4.66 万亿 tokens,约等于 3.7 万亿汉字,相当于 510 万本《红楼梦》的文本量。
这个调用量级的平台,肯定不是吸引一些人随便注册、用两下,就能跑出来的量级。
一定是有一批真实用户存在重度使用场景,把Agnes放进了日常使用里:拿它写代码、处理长文档、跑 Agent,批量生成图片和视频,这非常考验其稳定性和本身的模型能力。
把GitHub变看板
用透明换取工程信任
大规模的被使用,其实也是一种压测,毕竟免费开放之后,真正考验不是规模,而是稳定性。
人一多,问题自然也会跟着冒出来,比如响应变慢、图片生成失败、客户端接入报错等等,而Agnes比较加分的地方,是它直接把 GitHub Issues 和项目看板挂了出来:
GitHub Issues :https://github.com/AgnesAI-Labs/Agnes-AI/issues
项目看板:https://github.com/users/AgnesAI-Labs/projects/1
包括 TTFT、图片生成、Codex / 客户端接入、功能排期、开发进度这些问题,都可以在 GitHub 上看到对应反馈和跟进。
这个处理方式很有意思,Agens在面对问题反馈的时候,没有做“客服式解释”,反而能把问题本身变成了产品的一部分,让用户也有一种与团队一起来优化服务的感受。
透明、直接、开放,这在开发者场景里,反而更稳。当我对模型服务商有完全的了解的时候,我就可以非常放心和全面的判断,某个场景是否可以用这几个模型了。
模型体验如何?
目前免费给出的都是Agens的Flash版本模型,此类模型用于日常对话、生图都是没有问题的,就像大家常用的豆包,其实默认的应该都是flash模型,他胜在轻便、速度快,日常够用。
而且,这几款模型都各有特点,用对场景就特别能打,性价比极高。
Agnes-2.0-Flash百万上下文
复杂工程与长篇处理的利器
免费的模型里有Agnes-2.0-Flash,这个最核心的能力之一,是 Agnes-2.0-Flash 正式原生支持了 1M Token 的超长上下文窗口。
对于 AI 开发者和企业用户来说,免费模型1M 上下文,是能直接改变构建系统逻辑的重要程度。
有了 1M 的免费窗口,我们可以直接把一整本技术手册、复杂的长篇小说,或者整个项目的代码库、配置文件一股脑丢进去。让模型直接在全局视角下帮我们定位 Bug、梳理架构或进行多轮超长对话。
之前为了成本、为了避免超出上下文窗口,做的的文档切片都可以直接跨过了。这项能力无需修改现有代码,只要在 1M 范围内即可正常调用。
需要强调下的是,Agnes 官方也非常坦诚,为了在 API 全面免费开放期间保障绝大多数开发者的调用稳定性,高峰期会将上下文上限动态限制在 512K,这里需要做好工程处理。
除了知识类,能够量大好用,我还做了个简单的Coding,比如「飞机大战」小游戏开发,完成度蛮高,逻辑、页面、子弹、记分都没有问题,做个日常的小工具是能胜任的。
比如生成一个飞机大战的游戏:
比如「数学可视化」,创建一个交互式工具来可视化欧拉恒等式:
Agnes-Image-2.1-Flash
免费的4K 图像生成模型
除了文本模型,免费的模型中,Agnes-Image-2.1-Flash 也迎来了 4K 输出能力的正式上线,最高支持生成 4096×4096 的超高清图像,并且囊括了 16:9、9:16 等全尺寸比例。
我还试了一下生成4K和1K的对比:
提示词:清幽山间古村早市,石板路两侧老式木铺,摊位摆放青花瓷器皿与竹编箩筐,屋檐下悬挂一排纸灯笼,有几位老人整理货品,空中飘着淡淡炊烟,远景薄雾松林,电影写实摄影,繁复瓷纹、清晰竹编肌理、精致木雕纹路,柔和清晨自然光,微观细节锐利,无模糊发虚
1080P效果:
4K效果:
根据出师表的内容,生成的出师表封面:
虽然精细度和质感与一线的生图模型有差距,但是胜在性价比!但在“免费”加持下,它的性价比是无可匹敌的。作为免费的模型,Agnes 足够承担起视觉引擎的基础底座任务,极大降低了批量跑图的成本试错门槛。
能不能薅起来羊毛,就看各位的应用场景是否匹配了。
Agnes-Video-2.0
原生音画视频抽卡机
视频模型方面,Agnes-Video-2.0 目前已经支持原生音画同步生成,并提供了 720P 和 1080P 的输出选项。
支持文生视频、图生视频、多图视频生成以及关键帧动画,在实际使用中,我发现如果在单次 15s 的生成中加入过多的复杂动作指令,容易硬拼接、转场生硬、模型会出现严重的漂移,动作、场景的遵循能力低。不过参考图片生成视频时,视频中的人物与图片一致性保持还不错。
我体验下来,只用短时间的文/图生视频,一次尽量控制在1个场景中效果会比较稳定,类似于下边这样的:
三步嵌入现有工作流
前往Agnes的官网进行登录注册:https://platform.agnes-ai.com/
在API密钥中创建你的密钥:
根据文档(https://agnes-ai.com/doc),选择对应模型使用:
Agens AI还有一点好处就是,模型接口统一,在你原本的工作流中只需要替换一下基本的配置信息,就可以无缝接入。
写在最后
体验下来,Agnes 给我的最大感受是:它正在把“前沿大模型”变成真正的“开发者基建”。
过去我们在做系统架构设计,或者做一些新应用的快速构建时,往往要先在后台算一笔账:Token 消耗多少?长上下文要不要做切片?图片批量生成的成本有多高?这种成本焦虑,直接限制了技术落地的想象力。
但全模态免费的 Agnes 彻底打破了这个屏障。它提供了一个足够能打、且完全无需顾虑试错成本的底层底座。
以前需要精打细算的事情,现在可以大胆去跑、去尝试。当调用 AI 像接入水电、网络一样自然,不再需要时刻盯盘计费时,我们可以无视Toekn消耗,把精力完全转移到真正的“业务落地与场景创新”上。
这或许正是 Agnes 那句口号最打动人的地方:
让世界级 AI,属于每一个人。
我想,属于每一个人的,不仅是一个对话的AI接口,
更是放手去创造的自由。
我是梦飞,我们下次见~
往期推荐
Doubao 2.1 Pro深度实测:Agent 能力全面跃升
把 8 个节点压成 3 个后,我看懂了Step 3.7 Flash
Recently, Agnes AI has started appearing in discussions among developers and creators.
The reason is straightforward: this vendor has made its entire API completely free.
Text, image, and video - all modalities, all free.
Yes, no trial quotas, no tiered pricing, no thresholds or fuss - it is simply ready to use.
On the surface, this might look like "just another vendor doing a free strategy" - there are plenty of free options on the market, but very few that are actually good.
However, if you look at its call volume data over the past week, you will see that things are different.
4.66 trillion tokens in a single week
A real-world "stress test" scorecard
According to the latest official data from Agnes AI, Agnes's full-modality model has reached 4.66 trillion tokens in total weekly call volume:
Looking at 4.66T (trillion) alone, you might not have an intuitive sense of scale.
This call volume is second only to DeepSeek V4 Flash's weekly call volume on OpenRouter, which fully reflects its current scale of service.
Roughly converting 1 token to about 0.8 Chinese characters, 4.66 trillion tokens equals about 3.7 trillion Chinese characters, equivalent to the text volume of 5.1 million copies of Dream of the Red Chamber.
A platform at this call volume is not something that could be reached just by attracting a few people to casually register and try it a couple of times.
There must be a group of real users with heavy usage scenarios who have put Agnes into their daily workflow: using it to write code, process long documents, run agents, and batch-generate images and videos - this truly tests its stability and model capabilities.
Turning GitHub into a public dashboard
Using transparency to earn engineering trust
Large-scale usage is itself a form of stress testing. After all, once it is free and open, the real test is not scale, but stability.
When more people come, problems naturally surface - such as slower responses, image generation failures, client integration errors, and so on. Where Agnes earns extra points is that it directly publishes its GitHub Issues and project board:
GitHub Issues: https://github.com/AgnesAI-Labs/Agnes-AI/issues
Project board: https://github.com/users/AgnesAI-Labs/projects/1
Issues such as TTFT, image generation, Codex/client integration, feature scheduling, and development progress can all be seen with corresponding feedback and follow-up on GitHub.
This approach is quite interesting. When facing problem reports, Agnes does not give "customer-service-style explanations." Instead, it turns the problems themselves into part of the product, giving users a sense of optimizing the service together with the team.
Transparent, direct, and open - in developer scenarios, this actually feels more reliable. When I have full visibility into a model provider, I can make confident and comprehensive judgments about whether a given scenario can use these models.
How is the model experience?
Currently, what is being offered for free is Agnes's Flash version models. These models are perfectly fine for daily conversation and image generation. Like Doubao, which everyone commonly uses, its default is likely a Flash model - it wins on being lightweight, fast, and sufficient for daily use.
Moreover, these models each have their own characteristics. When matched to the right scenario, they perform exceptionally well and offer extremely high cost-performance.
Agnes-2.0-Flash with 1M context
A powerful tool for complex engineering and long-form text processing
Among the free models is Agnes-2.0-Flash. One of its core capabilities is that Agnes-2.0-Flash now natively supports an ultra-long context window of 1M tokens.
For AI developers and enterprise users, a free model with 1M context is significant enough to directly change how you structure your system logic.
With a free 1M context window, we can directly dump an entire technical manual, a complex novel, or the full codebase and configuration files of a project into it. The model can then help us locate bugs, sort out architecture, or conduct multi-round ultra-long conversations from a global perspective.
The document chunking we used to do for cost reasons or to avoid exceeding the context window can now be skipped entirely. This capability requires no changes to existing code - as long as it stays within 1M, it can be called normally.
It should be emphasized that Agnes is also quite candid: to ensure stable calling for the vast majority of developers during the period when the API is fully free, the context limit will be dynamically capped at 512K during peak hours. You should plan your engineering accordingly.
Beyond knowledge-type tasks that are high-volume and usable, I also did some simple coding - for example, developing a small "Airplane War" game. It had a fairly high level of completion; the logic, pages, bullets, and scoring all worked fine. It is capable of serving as an everyday small tool.
For example, generating an Airplane War game:
Followed
Follow
Replay Share Like
Close
Watch more
More
Exit fullscreen
Mengfei AI Followed
Share video
, duration 00:28
0/0
00:00/00:28
Switch to landscape mode
Continue playing
Progress bar, 0 percent
Play
00:00
00:28
00:28
Speed
Fullscreen
Speed playback
0.5x 0.75x 1.0x 1.5x 2.0x
HD Smooth
Continue watching
Full-modality API completely free! This may be the highest cost-performance AI foundation right now
Watch more
Repost
,
Full-modality API completely free! This may be the highest cost-performance AI foundation right now
Mengfei AI Followed
Share Like Favorite
Synced to Moments Write your comment
Video details
For example, "mathematical visualization" - creating an interactive tool to visualize Euler's identity:
Followed
Follow
Replay Share Like
Close
Watch more
More
Exit fullscreen
Mengfei AI Followed
Share video
, duration 00:11
0/0
00:00/00:11
Switch to landscape mode
Continue playing
Progress bar, 0 percent
Play
00:00
00:11
00:11
Speed
Fullscreen
Speed playback
0.5x 0.75x 1.0x 1.5x 2.0x
HD Smooth
Continue watching
Full-modality API completely free! This may be the highest cost-performance AI foundation right now
Watch more
Repost
,
Full-modality API completely free! This may be the highest cost-performance AI foundation right now
Mengfei AI Followed
Share Like Favorite
Synced to Moments Write your comment
Video details
Agnes-Image-2.1-Flash
A free 4K image generation model
In addition to the text model, among the free models, Agnes-Image-2.1-Flash has also officially launched its 4K output capability, supporting ultra-high-definition images up to 4096×4096, and covering all aspect ratios including 16:9 and 9:16.
I also tried comparing 4K and 1K generation:
Prompt: A quiet mountain village morning market at dawn, old wooden shops along the stone-paved road, stalls displaying blue-and-white porcelain vessels and woven bamboo baskets, a row of paper lanterns hanging under the eaves, a few elderly people arranging goods, faint cooking smoke drifting in the air, misty pine forest in the distance, cinematic realistic photography, intricate porcelain patterns, clear bamboo weave texture, exquisite wood carving details, soft natural morning light, sharp micro details, no blur or softness
1080P result:
4K result:
A cover for the "Chu Shi Biao" generated based on its content:
Although the level of detail and texture still lags behind top-tier image generation models, it wins on cost-performance! And with the "free" factor, its cost-performance is unmatched. As a free model, Agnes is fully capable of serving as the foundational visual engine, greatly lowering the cost and barrier for trial and error in batch image generation.
Whether you can take advantage of this free opportunity depends on whether your application scenarios match.
Agnes-Video-2.0
A native audio-visual video gacha machine
On the video model side, Agnes-Video-2.0 now supports native synchronized audio-visual generation and offers 720P and 1080P output options.
It supports text-to-video, image-to-video, multi-image video generation, and keyframe animation. In actual use, I found that if too many complex motion instructions are added to a single 15s generation, it tends to produce hard stitching, abrupt transitions, and severe model drift, with poor adherence to motion and scene. However, when generating video from a reference image, the character in the video maintains good consistency with the image.
From my experience, using short text/image-to-video and keeping each generation to roughly one scene yields more stable results, like the example below:
Followed
Follow
Replay Share Like
Close
Watch more
More
Exit fullscreen
Mengfei AI Followed
Share video
, duration 00:05
0/0
00:00/00:05
Switch to landscape mode
Continue playing
Progress bar, 0 percent
Play
00:00
00:05
00:05
Speed
Fullscreen
Speed playback
0.5x 0.75x 1.0x 1.5x 2.0x
HD Smooth
Continue watching
Full-modality API completely free! This may be the highest cost-performance AI foundation right now
Watch more
Repost
,
Full-modality API completely free! This may be the highest cost-performance AI foundation right now
Mengfei AI Followed
Share Like Favorite
Synced to Moments Write your comment
Video details
Three steps to integrate into an existing workflow
Go to Agnes's official website to log in and register: https://platform.agnes-ai.com/
Create your API key in the API keys section:
According to the documentation (https://agnes-ai.com/doc), select the corresponding model for use:
Another advantage of Agnes AI is that the model API is unified. In your existing workflow, you only need to replace the basic configuration information to connect seamlessly.
Final thoughts
After trying it out, my biggest impression of Agnes is this: it is turning "cutting-edge large models" into true "developer infrastructure."
In the past, when designing system architecture or rapidly building new applications, we often had to first do the math in the background: How many tokens will be consumed? Should long contexts be chunked? How high is the cost of batch image generation? This kind of cost anxiety directly limited the imagination for technical implementation.
But the fully free, full-modality Agnes has completely broken down this barrier. It provides a foundational layer that is capable enough and requires zero concern about trial-and-error costs.
Things that used to require careful budgeting can now be boldly run and experimented with. When calling AI becomes as natural as connecting to electricity, water, or the internet - without needing to constantly watch the meter - we can ignore token consumption and shift our full attention to real "business implementation and scenario innovation."
This is perhaps the most compelling part of Agnes's slogan:
Let world-class AI belong to everyone.
I think what belongs to everyone is not just an AI interface for conversation,
but also the freedom to let go and create.
I'm Mengfei, see you next time~
Previous recommendations
College application tutorial: save 10,000 yuan in consulting fees, the more you talk the more accurate it gets
Doubao 2.1 Pro in-depth review: Agent capabilities fully upgraded
Whose pockets did Mixue Bingcheng's money go into? MiniMax M3 dug through the whole internet
Yiwu AI store went viral; I used this tool to run a one-person company experiment
After compressing 8 nodes into 3, I understood Step 3.7 Flash












