Flash Models, Not Flagships, Are the Real Workhorses for Agent Workflows

Recently, GLM 5.2 has been making waves, and domestic models are buzzing again.

Along with DeepSeek V4, MiniMax M3, and Stepfun's Step-3.7-Flash, this wave of domestic large models is seeing fierce competition, and the heat has suddenly risen again.

Some friends might not be familiar with Stepfun models. Stepfun is also one of the AI Six Little Tigers.

For us AI bloggers, we use these models daily. The usage of these models generally falls into two categories.

'Pro/Flagship' and 'Flash/Efficiency' types

Pro/Flagship: Higher Capability Ceiling

This tier represents each company's strongest model, typically suitable for complex reasoning, long-chain planning, multi-turn task decomposition, code architecture design, and deep research scenarios.

Pro tier can be understood as the flagship capability layer of each company's models, mainly targeting complex reasoning, code engineering, long-chain Agents, and high-value tasks. Overseas representatives include the GPT flagship series, Claude Opus, Gemini Pro; domestically, it corresponds to the DeepSeek Pro series, Qwen Max, high-capability Kimi, Doubao flagship model, and GLM's high-capability versions.

The advantage of these models is high capability, high stability, and more stable understanding of complex tasks. But the cost is also obvious: higher cost, and speed is not necessarily the fastest.

Flash/Efficiency: The Balance Point of Model Capability

The Flash tier is more suitable for high-frequency calls in production environments.

It doesn't necessarily pursue first place on all benchmarks, but needs to do three things: fast response, low cost, and stable task completion rate.

In various Agent calls, such as data processing Agents, office Agents, etc., models that require continuity and cost control are needed.

It is not a 'low-spec Pro', but rather an independent category oriented towards efficiency-type Agent scenarios.

Let's test the actual effect.

The tool used is Trae. Global use of unified Trae settings, same project.

Each model is run individually. Before the test starts, the project environment and cache state remain consistent.

Create a project test set to check the model's capabilities in high-frequency tasks, code quality, speed, and other aspects.

Step-3.7-Flash

More suitable for placing into Agent workflows.

Previously made a news collection project, needed to develop some test classes to test the stability of the interface.

First, let AI organize the test prompts.

1. Test AI content recognition logic. Cover positive samples, negative samples, blacklist samples, short title samples.

2. Test event merging logic

3. Verify that the return structure is unified as:

• success
• data
• message

1. Test information source creation, update, enable/disable, verification failure scenarios.

2. Verify that the score changes are correct for official sources, multiple sources, within 24 hours, AI tags, high-priority sources.

Throw the prepared prompts to it first.

In about 5 minutes, the test class was generated.

After optimizing dependencies, the Agent automatically ran 42 tests.

About 80% was completed in one go. Due to some dependency issues, a second round of detection and modification was carried out.

The code volume is around 900 lines.

The total token consumption for input and output was around 5 million.

The cost was around 3.5 yuan.

Then we used the GPT5.4 model to run the same test.

GPT5.4

GPT5.4 belongs to the Pro tier model.

For the same task, the code quality written by GPT was relatively better, with no secondary optimization in the middle. For the code test set, the gap between the two was not large.

This type of test code is relatively simple, mainly looking at the model's multi-path writing ability.

GPT5.4's consumption was much higher than Step-3.7-Flash. Because GPT5.4 requires long thinking, for simple high-frequency tasks, the time might be relatively slow.

In Trae's resource management, you can see that this consumption was 1 dollar, which is about twice the usage of Step-3.7-Flash. Foreign models are inherently more expensive, Chinese models have a natural advantage.

The effect is about 90% of GPT5.4, the cost is 1/2 of GPT5.4.

Deepseek-V4-Flash.

Deepseek needs no introduction, it is the first stop for many people coming into contact with domestic large models.

Its biggest feature is not being flashy, but being stable, cheap, and having a mature ecosystem. Whether you are writing articles, revising copy, organizing materials, writing code, or making plans, it can handle most scenarios.

If other models are somewhat specialized in specific capabilities, Deepseek is more like a general-purpose base.

Here we use deepseek-V4-Flash.

Similarly, test the quality and timeliness of this model's generated test set code in Trae.

A few minutes later, the file was created.

There was a small problem, part of the code also had error messages, requiring secondary adjustment.

The generated code quality was acceptable, and the test chain could run through. However, there was one small issue: when testing the code written by deepseek-V4-Flash itself, the time consumed was longer than the previous two models, possibly because the test content was more.

It consumed about 1.2 million tokens, with a cost of about 0.2 yuan. There's no way around it, deepseek's price is indeed its biggest advantage.

GLM5.2

GLM5.2 belongs to the Pro tier model.

GLM5.2 is more suitable for observation in long tasks and AI Coding scenarios.

For this type of model, you can't just look at whether it can chat. More critically, it's about whether it can persist in running a relatively long task.

For example, let it read the entire project code, understand the directory structure, analyze where the problem is, and then modify, test, and continue fixing step by step. This process places very high demands on the model. If the context is insufficient, it will forget what it saw earlier; if tool calls are unstable, it will break halfway; if planning ability is poor, it's easy to go off track while modifying.

The positioning of GLM5.2 is very clear: long context, long tasks, Agent workflows.

Test the same task to see the effect. The total runtime was around 15 minutes, which is characteristic of Pro models. The capability is undeniably strong, but it inevitably lags behind in timeliness.

The required test code was also generated normally.

The Agent automatically ran tests for us. Multiple tests were conducted for each category.

There was one failure in the test. The Agent identified it and quickly repaired it.

But GLM5.2 has a very obvious problem, which is instability. The instability is because too many people are using it, sometimes requiring queuing, and secondly, there is the price issue.

This task consumed 9.8, but actually consumed about 12 yuan.

I also planned to try using an aggregation platform to test, using gemini-3.5-flash on the aggregation platform for testing. But halfway through, it failed because it was too expensive. I thought ten yuan could finish this test, but halfway through, it prompted insufficient balance.

Moreover, using an aggregation platform also has the problem of instability.

Finally

So after this test, my positioning of Step-3.7-Flash has become a bit clearer.

It's not about competing head-on with the Pro tier in extreme reasoning, nor about competing with the efficiency tier for the lowest unit price. It's more like being stuck in that most practical middle position: fast enough speed, controllable cost, and stability that can support continuous tasks.

Especially in production-grade Agent scenarios, this advantage will be more obvious. For example, high-frequency calls, multi-turn execution, low-latency response, code testing, data processing, office automation, plus some multimodal input. These types of tasks don't necessarily need the strongest model, but they definitely need a model that runs fast, runs stably, and runs cheaply.

From this perspective, Step-3.7-Flash is more like the comprehensive optimal solution in the 'efficiency frontier' track.

If your task is complex long-chain reasoning, deep research, or architecture-level code design, then it's still better to prioritize the Pro tier. But if it's a high-frequency Agent workflow in a daily production environment, I would be more inclined to put Step-3.7-Flash on the candidate list first.