Don't Be Fooled by Model Marketing — Real Agent Tasks Reveal the Truth in One Run

There are indeed more and more models available on the market now, each with its own highlights and focus areas. Just looking at promotional materials and benchmark scores makes it very hard to judge which one truly suits you — especially when tasks extend from single-turn conversations to multi-step operations, the situation becomes even more complex.

So I thought, why not pull out several mainstream models and actually run them through their paces to see how they really perform, and get a feel for them myself.

Evaluation Note: This is not a strict benchmark evaluation. It's more like an experiential observation record focused on a single long-chain Agent task. The results are for reference only and do not constitute a comprehensive verdict on the models.

The models used in this test include MiniMax-M3, DeepSeek-V4-flash, Step-3.7-flash, GLM5.2, and Gemini3.5 flash.

The test task was to create a single HTML page for an "AI Website Aggregation Platform." There were three core observation points:

1. Whether it could continuously call tools to complete the task

2. Whether it could stably generate a runnable page

3. Whether it would proactively check and fix issues after the page was completed

The WorkBuddy Agent tool was used globally. The cost is an estimate of the WorkBuddy platform's consumption for this task and does not represent official API pricing.

The general prompt is as follows:

Please complete the full development task of an 'AI Tool Navigation Site', requiring independent completion from requirement understanding to page generation, data organization, code implementation, run checks, and issue fixing.
Task Objective:
Create a complete, runnable single HTML page website with the theme 'AI Tool Navigation Site'. The page is used to display different types of AI tools, suitable for web demos, course materials, or official account long images.
Task Requirements:
1. Information Collection
Search the web and compile 20 mainstream AI tools, covering categories such as AI Writing, AI Programming, AI Images, AI Video, AI Search, AI Office, etc. Each tool needs to include: tool name, owning company, main purpose, target audience, and official website link.
2. Data Organization
Group the tools by category and organize them into structured data. Ensure information accuracy, avoid duplicate tools, and cover both domestic and international tools.
3. Page Design
Generate a clean, modern, tech-feel HTML page. The page needs to include a top title area, category filter area, tool card area, recommended tools area, comparison table area, and summary description area.
4. Interactive Features
The page needs to support filtering by tool category, keyword search, expanding tool card details, and a back-to-top button.
5. Code Implementation
Implement using a single HTML, CSS, and JavaScript file, without backend dependencies. Public CDN icon libraries or lightweight chart libraries can be used, but the page must be directly runnable.
6. Run and Check
After completion, please self-check the page for code errors, style misalignment, invalid buttons, missing links, filter failures, etc. If problems are found, please proactively fix them.
7. Output Result
Finally, output the complete runnable HTML file content, along with a brief explanation: what data sources were used, what modules the page contains, and what interactive features it has.
Special Requirement:
Please try to complete the entire task in one go. If tools such as search, web reading, code generation, file modification, run checks, error fixing, etc., are needed during the process, please complete them continuously in a reasonable order without skipping steps. The final result is based on a runnable page.

MiniMax-M3

MiniMax-M3 performed relatively stably in this type of long-chain task.

It basically proactively makes multiple rounds of tool calls, including searching for information, organizing data, generating page code, checking files, fixing issues, etc. The whole process is more like a normally functioning Agent, not just stopping at the level of "giving a piece of code."

During the test, the probability of tool call failure was very small, but not entirely zero. I had one tool call failure here, but it did not affect the final result generation. The model continued executing and produced the page.

This is the page effect after completion.

Judging from the final page, MiniMax-M3's data completeness, page structure, and interactive features are relatively complete. It doesn't particularly pursue visual flashiness, but its strength lies in process stability and relatively clear task understanding.

Points consumed in Workbuddy: about 27 points.

Converted, that's roughly 1.33 RMB.

If estimated by API unit price, MiniMax-M3 is a medium-to-low cost model, suitable for repeatedly running Agent workflow tasks.

After multiple tests, MiniMax-M3's task completion rate was 100% , and the tool success call rate was approximately 98% . A small number of tool calls failed, but did not affect the final result generation.

Simply put, MiniMax-M3's advantages are stability, low cost, and the ability to run to completion. It is suitable for batch page generation, data organization, code drafts, and lightweight Agent tasks.

Speaking of cheap, let's test the cheapest large model below — deepseek-v4-flash — to see how it performs.

DeepSeek-V4-flash

Using the same prompt, I also tested DeepSeek-V4-flash.

DeepSeek-V4-flash's overall speed is relatively fast, and its responses are very crisp. It performed well in understanding requirements, breaking down page modules, and generating HTML structures.

However, in long-chain tool calls, its style leans more towards "quickly completing the task." That is, it generates code very quickly, but in terms of data searching, data verification, and detail fixing, it is not as meticulous as MiniMax-M3 and Step-3.7-flash.

From the results, the page can be completed normally, and the basic modules are all there. For example, features like categories, cards, search, details, and tables are all covered.

DeepSeek-V4-flash is more suitable for speed-sensitive tasks. If you just want to quickly get a runnable HTML Demo, its efficiency is very high.

But if the task requires extensive data verification, page detail polishing, and multiple run-and-fix cycles, it sometimes needs a manual reminder. For example, asking it to check links again, optimize styles further, or supplement data fields.

Points consumed in Workbuddy: about 4 points. Converted, that's roughly 0.2 RMB.

From a price perception standpoint, DeepSeek-V4-flash's cost advantage is obvious, suitable for high-frequency calls.

After multiple tests, DeepSeek-V4-flash's task completion rate was approximately 100%. The tool call success rate was 99%.

My feeling is that DeepSeek-V4-flash is very suitable for a "quick generation + slight manual check" workflow. Speed and cost are good, but the detail stability of long-chain Agents still depends on the specific platform's tool environment.

Step-3.7-flash

Step-3.7-flash is the model in this test that most closely matches the positioning of a "production-grade Agent."

Its proactiveness in multi-tool calling is relatively high; it continuously completes searching, reading, organizing, generating, modifying, and checking. The whole process is more like fully executing a task, rather than simply answering a question.

The page effect is a typical dark tech style.

AI really likes this color scheme. Without specific instructions, many models will default to generating dark-themed website pages. This isn't necessarily bad, but if you want a clean, bright style leaning towards official account long images, it's best to specify it clearly in the prompt beforehand.

Step-3.7-flash performed quite prominently in data organization. The AI tool data is relatively complete, and the categorization is quite clear. It tries to cover different categories like writing, programming, images, video, search, and office, rather than just listing a few common tools.

From the perspective of page completeness, Step-3.7-flash has the highest content density. It tries to fill in all the modules required by the task, including the top title area, category filter, tool cards, recommended tools, comparison table, and summary description.

The test cost for this round was approximately: 0.7 RMB.

Judging from the unit price, Step-3.7-flash is a medium-to-low cost contender. Its advantage isn't in being the cheapest, but in "being able to run continuously, with few interruptions, and a high completion rate."

After multiple tests, Step-3.7-flash's task completion rate was approximately 100%, and the tool call success rate was approximately 99%.

If your task is high-frequency, multi-turn, low-latency, and includes tool chains like search, file operations, code, and fixes, Step-3.7-flash is a model worth putting on your candidate list.

GLM5.2

Next, let's look at the effect generated by GLM5.2.

GLM5.2 performed well in code generation and page structure. It can understand that this task requires a complete AI tool navigation site and can break down the page modules quite clearly.

From the results, the overall page completeness is acceptable. Content like categories, cards, search, and description areas are all covered.

GLM5.2's characteristic is relatively balanced capabilities. It can normally exert its model strength in Agent tasks; the biggest drawback is that it's too expensive.

The test cost for this round was approximately: 74 points. Converted, that's roughly 3.66 RMB.

Finally, let's test one more foreign model, Gemini3.5 flash, to see how it performs.

Gemini3.5 flash

If it's for frontend pages, Gemini's aesthetic sense has always been quite on point. So here I used the Gemini3.5 flash model.

Below is the AI tool navigation webpage effect produced.

Gemini3.5 flash's biggest advantage is comfortable page viewing.

The frontend pages it generates are more refined, with more comfortable layouts, and better use of whitespace and layering. Compared to the previous models, Gemini3.5 flash understands frontend design a bit better.

However, Gemini3.5 flash also has obvious problems.

It is indeed better in visual performance, but its data collection is not as extensive as the previous models. Especially compared to Step-3.7-flash, Step collected more complete data, had more comprehensive category coverage, and was more proactive in tool calling.

The test cost for this round was approximately: 9 RMB.

Gemini3.5 flash's price is significantly more expensive, especially for tasks with more output tokens, tool calls, and code generation. The cost will be much higher than domestic Flash-tier models.

If you have high requirements for page quality, you can try Gemini3.5 flash. It is suitable for display pages, official website demos, product introduction pages, and course material pages. But if you care more about cost and high-frequency calls, you still need to be cautious.

Test Result Comparison

Model	Task Completion Rate	Tool Call Success Rate	This Round Cost	Main Advantages	Main Disadvantages
MiniMax-M3	100%	~98%	~1.33 RMB	Stable, low cost, can run the complete process	Page aesthetics are mediocre, average visual impact
DeepSeek-V4-flash	100%	~99%	~0.2 RMB	Fast, low cost, suitable for quick first drafts	Detail checking and page polishing sometimes need manual reminders
Step-3.7-flash	100%	~99%	~0.7 RMB	Proactive tool calling, complete data coverage, strong long-chain execution feel	Page easily defaults to dark tech style, needs style constraints upfront
GLM5.2	100%	~97%	~3.66 RMB	Balanced overall capability, good page structure and code completion	Proactive searching, verification, and fixing execution feel is not the strongest
Gemini3.5 flash	100%	~96%	~9 RMB	Best page aesthetics, more mature layout, whitespace, and visual hierarchy	Significantly higher cost, data collection and tool calling proactiveness not as good as Step-3.7-flash

Summary

In this test, what I focused on more was not single-turn answering ability, but whether the model could run a real task from start to finish.

If only looking at page aesthetics, Gemini3.5 flash is indeed stronger. The web pages it generates look more like a finished demo, visually more comfortable.

If looking at tool calling and data completeness, Step-3.7-flash's performance is more prominent. It more proactively searches, organizes, generates, and checks, suitable for long-chain Agent tasks.

If looking at cost and stability, MiniMax-M3 is a very stable choice. It's not particularly flashy, but it can complete tasks in multiple tests, and tool call failures won't significantly affect the results.

DeepSeek-V4-flash's advantage is speed and low cost, suitable for quickly generating first drafts. GLM5.2 is more balanced, suitable for comprehensive tasks.

So model selection still depends on the scenario.

For display-oriented pages, prioritize Gemini. For production-grade Agent workflows, focus on Step-3.7-flash. For high-frequency, low-cost tasks, consider MiniMax-M3 and DeepSeek-V4-flash.