大家好，我是程序员海军，OPC｜AI 爱好者｜独立开发者。
GPT 算是我这两年多用得最勤的 AI 工具了。
从一开始拿它查资料、写代码，到后来做需求分析、开发 AI 应用，甚至平时一些琐碎的事情，我基本都是先打开 GPT。说它是我的主战场，一点也不夸张。
但最近几个月，用起来是真的有点烦。
有时候需求已经讲得很清楚了，它还是理解偏了；复杂一点的任务，来回沟通几轮也收不了尾。以前觉得它挺省心，现在反倒需要自己一直盯着、纠正。偶尔切到 Astra Pro 这种高强度模式，体验才稍微好一点。
当然，这是我自己的使用感受。我也不敢说一定是模型变差了，可能跟任务复杂度、使用方式、预期变化都有关系。但有一点我越来越确定：不能把所有活都押在一个模型上。
于是最近我重新研究了一圈模型 API 平台，中间也看过不少中转站。折腾完之后，最后选择了 Aiionly。正好手里有个一直想做的求职方向小项目，就拿它作为这次的应用场景。
为了换个模型，我又研究了一圈 API 平台
以前只是聊天，哪个工具顺手就用哪个，倒没觉得有什么问题。但自己开始做 AI 应用以后，情况就不一样了。
一款应用可能要处理长文本、识别截图、分析用户输入，还要连续对话。不同任务对模型的要求不一样，价格、响应速度和调用成本也得考虑。总不能每次想换个模型，就重新注册账号、充值、改一套接入逻辑。
这次我筛选平台，主要看三件事：模型选择够不够灵活，API 接入麻不麻烦，以及费用能不能看得清楚。
最后选的 Aiionly，定位就是一个多模型 API 聚合平台。简单理解，不用分别去不同模型厂商那里折腾，通过一个平台就能选择和调用多种模型。对我这种平时既写代码、又折腾 Agent 和各种小产品的人来说，这种方式要方便不少。
我比较关注的是它的统一 API 接口。对开发者来说，接口方式相对统一，后面需要换模型时，调整模型配置比重新接一套服务省事。当然，模型之间的参数、能力和返回格式仍然可能有差异，该适配的地方还是得适配。
另外就是费用。Aiionly 有按量调用，也提供 Token Plan 套餐，可以根据自己是偶尔测试，还是经常跑 Coding、Agent 任务来选。控制台能查看用量和账单，这一点我挺在意。做 Demo 的时候可能觉得几个 Token 无所谓，真要长期跑起来，成本就不能稀里糊涂了。
这里也说清楚，API 平台不是给模型加智商的。同一个底层模型，不会因为换个平台就突然聪明很多。
我看重的是模型选择权和接入便利，而不是某种所谓的“神奇加成”。至于长期稳定性、实际速度这些，还是要放到自己的项目里持续验证，不能只看宣传页面。
AiiOnly 和 WorkBuddy 已完成模型路由打通。现在也可以在 AiiOnly 客户端进入「Agent 路由 → WorkBuddy」，即可一键选择所需模型并完成配置。
使用也很简单，它提供了有个OpenAI 兼容接口，只需要在 Workbuddy 或者其它AI 工具(Trae, Codex)配置也可以使用，这里我们拿 Workbuddy 演示操作：
今年面了 50 多个人，我想做个求职避坑 Agent
说到这个小项目，起因其实挺简单的。
今年我陆续面试了 50 多位候选人，平时也会有朋友、网友来找我聊简历、面试和找工作的事情。接触得多了，我发现很多人找工作时，关注的往往是“我能不能通过面试”，但还有一个问题容易被忽略：这家公司、这个岗位，到底值不值得我花时间去？
去年我自己也碰到过一次工作机会和实际情况差距很大的事。虽然最后及时止损了，但中间花掉的时间和精力，已经收不回来了。也因为这个经历，我对“先了解清楚，再做选择”这件事挺有感触。
尤其现在找工作也不容易。投简历、约面试、准备笔试，有时候还得专门请假跑一趟。聊到最后才发现薪资口径不一致、工作时间接受不了，或者岗位职责和最初看到的完全不是一回事，时间就这么搭进去了。
于是我开始琢磨，能不能做一个 Agent，先帮求职者把招聘信息过一遍。
最早想得挺直接：识别垃圾岗位、僵尸岗位、坑人岗位。但真正梳理需求时发现，这个说法还是太粗了。光凭一张招聘截图，AI 很难判断一家公司到底靠不靠谱；岗位写得模糊，也不能直接说人家是骗子。
后来我把方向改成了一个更实际的东西，暂时叫它「职场评测」。它不负责给公司判好坏，而是结合求职者自己的情况，判断这个机会要不要继续投入时间。
我的思路是，用户先上传简历，设置自己的求职偏好，比如期望城市、薪资、双休、远程办公这些条件。接着把招聘 JD、公司介绍、HR 的聊天截图丢进去，让 AI 辅助分析。
这里有三个问题要分开看：一是你有没有能力胜任；二是这个岗位是不是符合你的个人要求；三是招聘信息里有没有需要警惕或者进一步核实的地方。
举个例子，一个岗位写着「20～30K，包含绩效」，而你的要求是固定月薪至少 20K。哪怕技能匹配度再高，也不能直接告诉你“符合要求”。真正该做的是把薪资结构标成待核实，再告诉你下一步应该怎么向 HR 问清楚。
再比如，招聘信息里没有提加班情况，不代表公司一定双休；也不能因为没写，就认定它是大小周。不知道就是不知道，不能让 AI 硬猜。
最后
这个项目其实也让我重新想了下，自己到底需要怎样的 AI 工具。
现在模型更新太快了，今天这个强，过段时间可能又换一个。每次新模型发布，大家都忙着测榜单、比较谁更聪明。我也会关注这些，但真到做项目的时候，最重要的还是它能不能把手上的任务完成。
比如这个求职 Agent，截图理解、简历提取、岗位分析可能会用到不同的能力。后面哪个模型处理得更稳、成本更合适，就让它负责对应任务，而不是从头到尾只能用一个模型。这也是我选择 Aiionly 这类多模型入口的原因。
当然，平台只是其中一环。AI 分析是否靠谱，还得看需求有没有定义清楚、输入材料是否充分、结果有没有证据支撑。尤其涉及求职决策，不能把 AI 的推测包装成事实。简历和 HR 聊天记录里有个人信息，使用时也得注意打码和隐私保护。
我现在依然会用 GPT，它毕竟陪我干了不少活。只是相比以前，我不太想再把自己绑在某一个工具或者模型上了。工具可以换，项目和解决问题的能力，才是自己真正能留下来的东西。
最后放一下这次用到的平台，感兴趣的朋友可以自己看看：
Aiionly 官网： www.aiionly.com/
我的邀请注册链接： maas.aiionly.com/login?invit…
目前的新人活动里，通过这个链接注册，可以领取一张 16-15 的平台金币充值代金券，另外还有一张 5 元无门槛 Token Plan 代金券。具体领取条件、有效期和适用范围，以注册后的活动页面为准。
如果平时只是偶尔问几个问题，其实没必要专门折腾 API。但如果你跟我一样，经常用 AI 编程、做 Agent，或者想把某个自己的想法做成产品，多模型接入这种方式还是值得了解一下。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Hello, I'm a programmer, Navy, OPC|AI lover, independent developer.
GPT is kind of the first AI tool I've used in over two years.
From the beginning of the search, the writing of the code, to the subsequent needs analysis, the development of the AI application, and even the usual triviality, I basically opened the GPT. Saying it is my battleground, it is no exaggeration.
But in recent months, it's really annoying.
Sometimes the needs have been made clear, and it has been understood to be biased; a more complex task cannot end with a few rounds of communication. It was thought to be a relief, but now it needs to keep staring at and correcting itself. It's only a little better to cut into this high-strength pattern of Astra Pro.
Of course, that's how I used it. And I can't tell you that the model has changed, and it may have something to do with the complexity of the task, the way it is used, the changes that are expected. But there's one thing I'm increasingly sure of: I can't put everything on a model.
So I've recently re-examined a model API platform, and I've seen a lot of transit stations in the middle. After that, the final choice was "Aiionly." I've got a little job-seeking project that I've been trying to do, and I'm using it as an application.
I've been studying the API platform for a different model.
I used to just talk. I didn't think it was a problem. But when I started doing AI applications, it was different.
An application may involve processing long text, identifying screenshots, analysing user input and continuous dialogue. The requirements of the model vary from mission to mission, and prices, response speed and call costs are considered. It is impossible to re-register the account number, top up and change the access logic every time you want to change the model.
This time I screened the platform and looked at three things: whether model selection is flexible enough, API access is not a problem, and whether the cost is clear.
The final selection, Aimonly, is a multi-model API assembly platform. A simple understanding is that multiple models can be selected and called upon through a platform without having to go to different model manufacturers separately. It's a convenient way for someone like me who usually writes both code and agent and a variety of small products.
I'm more concerned with its unified API interface. For developers, the interface approach is relatively uniform, and it is cheaper to adjust the configuration of the model than to re-establish a set of services when the following model is needed. Of course, there may still be differences between the parameters, capabilities and return formats of the model, and the suitable place is still appropriate.
And the cost. Aiionly has a volume call and offers a Token Plan package, which can be chosen from an occasional test or from a regular running of Coding, Agent missions. The console has access to volumes and bills, which I care about. When you're doing Demo, you probably think a few Tokens don't matter. If you really have to run for a long time, you can't be so confused.
It's clear here that the API platform is not an IQ for models. The same bottom model doesn't suddenly get smarter because of a different platform.
I look at model selection and accessibility rather than some sort of "magic addition." As for long-term stability and real speed, they should be continuously validated in their own projects and should not be viewed only on promotional pages.
AiiOnly and WorkBuddy have completed model route access. Now you can also enter "Agent Route →WorkBuddy" on the AiiOnly client, select the desired model and complete the configuration.
It's also easy to use, and it provides an Open AI compatible interface that can only be used in the WorldBuddy or other AI tool configuration, where we take the Workbuddy demonstration:
It's 50 people this year. I'd like to do a job, Agent.
And speaking of this small project, it's quite simple.
I've been interviewing over 50 candidates this year, and I usually have friends and netizens coming to me to talk about résumés, interviews and job search. Much more, I find that a lot of people looking for jobs tend to focus on “can I pass the interview”, but there's one more question that can be ignored: is this company, this job, worth my time?
Last year, I myself experienced a large gap between jobs and reality. Although the damage was eventually stopped in time, the time and effort spent in the middle were not recovered. And because of that experience, I'm touched by the fact that “know first and choose”.
Especially now it is not easy to find a job. Biographical information, interviews, preparation of written examinations and sometimes special leave to travel. At the end of the discussion, it became clear that pay was not consistent, that working hours were not acceptable, or that the job responsibilities were not the same as what was initially seen, and that the time was just in.
So I began to wonder if I could be an Agent and help the job-seekers run the job.
First thought was straightforward: identification of garbage jobs, zombie jobs, pithead jobs. But when it comes to really combing needs, it turns out that it's still too rough. It is difficult for AI to judge whether a company is on its own or not; it is vaguely written, and it is not possible to speak directly of a fraud.
Then I changed my direction to something more practical and called it a "works assessment" for the time being. It is not responsible for making good or bad decisions for companies, but rather for determining whether this opportunity should continue to be invested in the job-seekers themselves.
My idea is that users upload curricula vitae and set their own job preferences, such as expectations of cities, salaries, double breaks and teleworking. Then, the JD, the company's introduction, HR's chat clips were thrown in to allow AI to support the analysis.
There are three questions to be examined separately: first, whether you are competent; second, whether the post meets your personal requirements; and third, whether there is any need for vigilance or further verification in the recruitment information.
For example, a job says "20-30K, including performance", and you ask for a fixed monthly salary of at least 20K. Even a high skill match cannot be a direct “compatibility” to you. What really needs to be done is to mark the salary structure for verification and then tell you what next step should do to HR.
For example, the fact that overtime is not mentioned in the recruitment information does not mean that the company must take a double break; nor can it be assumed that it is a small week because it is not written. I don't know. I just don't know.
Finally.
This project actually rethought what kind of AI tools I needed.
Now the model is being updated too quickly, and today it's strong, and it might be another one later. Every time a new model, publish, everyone's busy with the list and who's smarter. I would also be concerned about this, but the most important thing when it comes to projects is whether it can do the task on hand.
This job search, for example, Agent, can be used to understand, extract and analyse jobs. The next model, which is handled more steadily and at a more appropriate cost, is assigned to the corresponding task, rather than using only one model from scratch. And that's why I chose such multi-model portals.
Of course, the platform is just one of them. AI analyses whether the demand is well defined, whether the input material is sufficient and whether the results are supported by evidence. In particular, decision-making about finding a job is not possible to package AI assumptions. Personal information is included in the curriculum vitae and HR chat records, which are also used with regard to codekeeping and privacy protection.
I'm still using GPT, and it's doing a lot of work with me. I just don't want to tie myself to a tool or model any more than before. Tools can be changed, projects and problem-solving capabilities are what they can really stay.
Finally, let's put on the platform that we used this time.
Aiionly Network: www.aiionly.com/
My invitation to register is: maas.aiionally.com/login?invit...
In the current event, a 16-15 platform gold top up voucher is available for registration through this link, along with a $5 token Plan voucher. Specific terms of receipt, duration and scope of application are based on the registered activity page.
If you ask only a few questions once in a while, there's no need to focus on API. But if you're like me, often using AI programming, doing Agent, or trying to make a product of one's own ideas, multimodel access is worth understanding.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.