我一直觉得,AI 烧钱这个事,跟我没关系。
那是大厂的事。是投资人的事。是 Sam Altman 半夜睡不着觉的事。
我?我就是一个蹭免费的。哪家不收钱我用哪家。Claude 用一阵免费额度,GPT 用一阵 free tier,谁开了新模型我就溜达过去薅一把。
说实话我没啥忠诚度。
但 Agnes 这个事,让我有点上头了。
3 万亿 Token,什么概念?
我算了一下,假设你一天写一万字的 prompt,每字大概折算成 4 个 token,就是一天 4 万个。3 万亿除以 4 万,等于……我懒得算了。反正就是好几千年。
但 Agnes 把它白送给用户了——从上线到现在,三周不到,全平台跑掉了 3.12 万亿个 token。
不是免费试用,是无限期、不限量、全球可用的全免费。
你说这公司图啥?我也不知道。但我觉得这事儿值得掰扯一下。
三个模型,全部免费
Agnes 现在挂出来的东西很简单:
| 模型 | 用途 | 亮点 |
|---|---|---|
| Agnes-2.0-Flash | 文本对话 | 百万上下文窗口 |
| Agnes-Image-2.1-Flash | 文生图 | 支持 4K 输出(灰度中) |
| Agnes-Video-V2.0 | 视频生成 | 也是免费的 |
没有分级套餐,没有 "每天免费 50 次,之后按量计费"。
就是打开就能用,用多久都行。API 和网页版同步免费——对,API 也是免费的。
你要拿它搭个 bot、写个自动回复、跑个批量任务,通通不收钱。OpenAI 兼容 API,代码一句话都不用改。
百万上下文不是噱头
现在大家都在卷上下文长度。动不动几百万 token,但真正能稳定用到 100K 以上的不多。
Agnes-2.0-Flash 这个百万上下文,我实测了一下——把一篇 4.6 万字的长文档丢进去,让它从头复述细节,再问中间的冷门信息点,基本都对得上。
上下文占用才 18%。
说实话,对大多数人来说,10 万 token 的上下文就已经够用了。百万上下文真正的价值不是让你一次塞一百万进去——而是你不用再焦虑 "长度够不够"。
你写 prompt 的时候心态是不一样的。敢给更多上下文,敢让模型记住更多细节,不用反复丢 "前面说的你还记得吗"。
那种 "不用省着用" 的感觉,确实挺爽的。
4K 生图也在路上了
图像这边也悄悄做了升级。
Agnes-Image-2.1-Flash 现在可以输出 4096×4096 的超高清了。我试了一下,prompt 完全不变,就改了 size 参数,从 1024×1024 换成 4096×4096,出来的图细节多了一截。
不过这个功能还在灰度,不是所有人都能用到。什么时候全量开放,没说。
但至少方向是好的——不是卡你分辨率逼你充钱,而是反过来:给你更高的上限,不收钱。
新加坡来的 "鲶鱼"
Agnes 背后的公司叫 Sapiens AI,新加坡的。
说实话我不太关心是哪国的。我只关心一个事:它能不能活下去。
免费是最难持续的商业模式。没有之一。
3 万亿 token 的算力成本,就算自己搭了推理集群,也不是个小数目。Sapiens AI 肯定有自己的算力优势或者某种补贴——但具体是什么,没人说得清。
作为用户,我的态度很明确:能用一天算一天。免费的时候使劲用,哪天它撑不住了也不意外。
毕竟,上一个这么玩的团队,现在已经成了 OpenAI 的对手盘。
说句实话,我现在写文章、写代码、生配图,都走 Agnes 了。不是因为它是最好用的——而是因为它不收钱,我不心疼。不心疼的时候,反而用得最多。
Agnes 有没有自己的问题?当然有。模型精度不如 GPT-4o 和 Claude 4,中文生图的稳定性有时候飘一下。
但你要知道——
它是免费的。
免费的,能稳定跑一个月,能出百万上下文,能出 4K 生图。
这不是技术问题,这是心态问题。
你愿意花多少钱跟 AI 对话,决定了你愿意拿 AI 做什么事。免费的,你什么都敢试。收费的,你每次都在算 "这值不值"。
Agnes 把 "值不值" 那个问题,直接帮你删掉了。
这是我最近最愿意推荐给身边人的一个产品。
不是因为它最牛。是因为它让我觉得:原来免费的,也能这么好用。
I've always felt that the burning money side of AI has nothing to do with me.
That's the big players' problem. It's the investors' headache. It's what keeps Sam Altman up at night.
Me? I'm just a free-rider. I use whoever isn't charging me. I get my free quota with Claude, my free tier with GPT, and I hop over to try out new models as they launch.
Honestly, I have no loyalty.
But Agnes has got me feeling something.
What does 3 trillion tokens mean?
I did a quick calculation. If you write 10,000 words a day as a prompt — roughly 4 tokens per character — that's 40,000 tokens a day. 3 trillion divided by 40,000 is... well, I'm too lazy to calculate, but it's several thousand years.
And Agnes gave all that away for free — from launch to now, less than three weeks, the whole platform ran through 3.12 trillion tokens.
Not a free trial, but permanently free, unlimited, and globally accessible.
What's their game? I don't know. But I think it's worth talking about.
Three models, all free
Agnes is very straightforward right now:
No tiered pricing, no "50 free credits per day, then you pay."
You open it up and start using it. Use it for as long as you want. Both the API and the web version are free — yes, even the API.
Build a bot, write autoreplys, run batch jobs — none of that costs money. And the API is OpenAI-compatible, so no need to change your code.
The million-context isn't just hype
Everyone is bragging about long context now. Millions of tokens? But few can actually handle more than 100K reliably.
I tested Agnes-2.0-Flash with a 46,000-word document. I asked it to summarize, to find specific details, and it held up perfectly.
Context usage was only 18%.
To be honest, for most people, 100K context is already plenty. The real value of a million-token context isn't that you'll use a million tokens at once — it's that you stop worrying about "running out of space."
It changes how you write prompts. You can load in more background, make the model remember more, without worrying if it'll forget something.
That "use it as much as you want" feeling? It's refreshing.
4K image generation is coming too
They're quietly working on the image side, too.
Agnes-Image-2.1-Flash now supports 4096×4096 resolution. I tried it — kept the prompt and style the same, just changed the size parameter from 1024×1024 to 4096×4096. The result had much more detail.
This feature is still in beta, not everyone can use it yet. No word on when it'll go public.
But the direction is right — they're giving us higher limits, not charging us for them.
The "disruptor" from Singapore
The company behind Agnes is Sapiens AI, based in Singapore.
I'm not very patriotic. I don't care where they're from. I only care about one thing: can they stay alive?
Free is the hardest business model to sustain. There is no exception to this rule.
3 trillion tokens of compute power, even if you built your own cluster, is a massive expense. Sapiens AI must have some kind of subsidy or cost advantage — but no one really knows what it is.
As a user, my stance is clear: I use it as long as I can. If it goes under, I'm not surprised.
After all, the last company that tried something like this ended up becoming OpenAI's rival.
To be honest, I do all my writing, coding, and image generation with Agnes now. Not because it's the very best — but because it's free, and I'm not stressing over tokens. When you're not stressing, you use it more.
Does Agnes have flaws? Sure. The model isn't as accurate as GPT-4o or Claude 4, and the image generation for Chinese text isn't always stable.
But remember —
It's free.
Free, with a stable month of operation, million-token context, and 4K image output.
This isn't a tech problem, it's a mentality problem.
How much you're willing to pay for AI determines what you're willing to do with it. Free means you try everything. Paid means you're always calculating if it's worth it.
Agnes removes that calculation for you.
That's why I recommend it to everyone.
Not because it's the strongest. But because it proves that free can be good enough.