跪拜 Guibai
← All articles
AI Programming · OpenAI · AIGC

Five AI Models Run the Same Agent Task: What Actually Ships vs. What Just Looks Good

By 猫猫头啊 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Model selection for agent workflows is still a guessing game when all you have are benchmark scores and marketing pages. This comparison surfaces the real trade-offs — cost, tool-calling persistence, data completeness, and visual quality — that determine whether a model actually finishes the job or just generates plausible-looking code.

Summary

Five models were given an identical prompt to build a full AI tool navigation site: search for 20 tools, organize structured data, generate a single-file HTML page with filtering and search, then self-check and fix issues. Every model hit a 100% task completion rate across multiple runs, but the path to that result varied significantly. MiniMax-M3 ran the most stable end-to-end process at roughly ¥1.33 per run, while DeepSeek-V4-flash delivered a working page for just ¥0.20, though it skipped some detail checks and needed occasional human nudging.

Step-3.7-flash stood out for data completeness and proactive tool calling — it searched, read, organized, generated, and checked without being prompted — making it the closest to a production-grade agent in the lineup. Gemini 3.5 Flash produced the most polished, visually refined page but collected less data and cost ¥9 per run, roughly 45 times more than DeepSeek-V4-flash. GLM5.2 landed in the middle: balanced capability at ¥3.66, but its tool-calling initiative lagged behind Step-3.7-flash.

The takeaway is not which model wins, but that the choice depends entirely on the job. A quick demo needs speed and low cost; a production agent pipeline needs thoroughness and autonomous error recovery; a client-facing showcase needs visual design sense. Benchmarks that only measure single-turn accuracy miss these trade-offs.

Takeaways
All five models achieved a 100% task completion rate on the AI tool directory task, but the quality, cost, and autonomy of execution differed substantially.
DeepSeek-V4-flash was the cheapest at ¥0.20 per run and the fastest, but it required occasional manual reminders for detail checks and style polishing.
MiniMax-M3 offered the most stable end-to-end execution at ¥1.33 per run, with tool-call failures that did not derail the final result.
Step-3.7-flash showed the strongest autonomous tool-calling behavior — searching, reading, organizing, generating, and checking without being prompted — at ¥0.70 per run.
Gemini 3.5 Flash produced the most visually polished page but collected less tool data and cost ¥9 per run, roughly 45× more than DeepSeek-V4-flash.
GLM5.2 delivered balanced capability at ¥3.66 but lagged in proactive searching, verification, and self-repair compared to Step-3.7-flash.
Most models defaulted to a dark tech-style color scheme unless the prompt explicitly requested a lighter, cleaner design.
Cost differences are extreme: the most expensive run (Gemini 3.5 Flash) cost 45× more than the cheapest (DeepSeek-V4-flash) for the same task.
Conclusions

Benchmark scores and single-turn accuracy tell you almost nothing about how a model behaves in a multi-step agent loop. The models that looked best on paper were not always the ones that ran the most complete process.

Tool-calling proactiveness is a separate axis from code-generation quality. Step-3.7-flash searched and verified data autonomously; DeepSeek-V4-flash generated code fast but skipped verification steps unless prompted.

Visual design quality and data thoroughness appear to trade off in current models. Gemini 3.5 Flash made prettier pages but collected less data; Step-3.7-flash gathered more complete data but defaulted to a generic dark theme.

Cost-per-task varies by a factor of 45× across these models for the same prompt, which makes model choice an economic decision as much as a technical one when running agent workflows at scale.

Every model completed the task, which suggests the floor for agent capability is rising fast — but the ceiling on thoroughness, autonomy, and polish still separates the contenders.

Concepts & terms
Agent workflow
A multi-step task where an AI model autonomously calls tools (search, code generation, file operations, error checking) in sequence to complete a goal, rather than answering a single prompt.
Tool calling
The ability of an AI model to invoke external functions — such as web search, file reading, or code execution — as part of its response, rather than relying solely on its training data.
Long-chain task
A task requiring many sequential steps and tool invocations, where the model must maintain context and make decisions across an extended interaction rather than producing a single output.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗