跪拜 Guibai
← All articles
AI Programming · AIGC · OpenAI

Agent Benchmarks Don't Matter — Completion Rate, Speed, and Cost Do

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

Agent workflows amplify small reliability gaps into broken pipelines. A model that disconnects once or picks the wrong tool doesn't just produce a worse slide — it breaks the chain and forces a restart. For teams building automated multi-step agents, completion rate and per-task cost matter more than any benchmark number, and Chinese Flash-tier models are hitting a price-performance point that GPT and Gemini aren't matching yet.

Summary

Five models were given identical prompts inside the Trae Work Agent framework: research short-video platforms and produce a PPT, then scrape Xiaohongshu for trending posts and compile an HTML report. Step 3.7 Flash completed both tasks reliably in about 5 minutes for roughly 1 yuan, with transparent tool-call logging. DeepSeek V4-Pro matched that speed at half the cost and produced better-organized content. MiniMax M3 ignored the PPT tool and generated an HTML presentation instead — visually decent but the wrong format — and handled scraping via internal search rather than opening a browser. Gemini 3.5 took 12 minutes, disconnected mid-task, and cost 6 yuan. GPT 5.4 ran 10 minutes, cost 9.5 yuan, and delivered mediocre output after an initial tool-selection mistake.

Information scraping exposed deeper differences. Step 3.7 Flash burned nearly 2 million tokens navigating Xiaohongshu's DOM to extract likes and comments correctly. DeepSeek V4-Pro used 3.6 million tokens but cost only 0.5 yuan. MiniMax M3 sidestepped the browser entirely and relied on internal search, producing acceptable but differently-sourced results.

The takeaway is not that one model wins. Flash-tier models carve out a distinct niche: they trade benchmark-topping scores for the speed, cost, and completion reliability that production Agent pipelines actually demand.

Takeaways
Step 3.7 Flash completed both PPT generation and browser scraping reliably, with clear tool-call logs, in about 5 minutes for ~1 yuan.
DeepSeek V4-Pro matched Step's speed, cost half as much (~0.5 yuan), and produced better-organized content with more vibrant slides.
MiniMax M3 ignored the PPT tool and generated an HTML presentation instead; it also handled scraping via internal search rather than opening a browser.
Gemini 3.5 disconnected mid-task, took 12 minutes, and cost 6 yuan — its visual aesthetics don't compensate for unstable Agent chains.
GPT 5.4 initially produced HTML instead of PPT, then delivered mediocre results after switching tools, costing 9.5 yuan over 10 minutes.
Browser-based scraping burned 2–3.6 million tokens per task; Step 3.7 Flash correctly parsed DOM elements, while MiniMax M3 bypassed the browser entirely.
Flash-tier models prioritize speed, cost, and completion stability over leaderboard scores, which aligns with what production Agent pipelines actually need.
Conclusions

Leaderboard rankings and launch-event demos measure a model's peak capability on isolated prompts; Agent tasks measure whether a model can chain multiple tool calls without derailing — and the two don't correlate well.

Tool-selection errors are a silent failure mode. MiniMax M3 and GPT 5.4 both produced usable output in the wrong format, which in a production pipeline means a downstream system breaks or a human has to intervene.

Cost differences compound fast: running GPT 5.4 at 9.5 yuan per task versus DeepSeek V4-Pro at 0.5 yuan means a 19x cost multiplier for worse results, which is unsustainable at any scale.

Gemini's instability inside a third-party Agent framework suggests that model quality is partly a function of the runtime environment; a model that runs well in a first-party sandbox may fall apart when orchestrated by external tooling.

MiniMax M3's decision to skip the browser and use internal search for scraping is a pragmatic shortcut that produced acceptable results — it raises the question of whether full browser automation is always worth the token cost.

Concepts & terms
Agent task
A compound AI workflow where a large language model autonomously chains multiple steps — understanding a prompt, searching the web, reading pages, extracting information, generating code, and calling external tools — to produce a final deliverable like a PPT file or a data report.
Flash-tier model
A class of LLMs optimized for speed and cost rather than maximum reasoning depth or benchmark scores. They target high-frequency, production-grade Agent pipelines where completion reliability and per-task cost outweigh single-output quality.
Tool invocation / tool call
The mechanism by which an LLM Agent selects and executes an external function — such as a PPT generator, browser, or code interpreter — as part of a multi-step task. Incorrect tool selection is a common failure mode in Agent workflows.
End-to-end Agent chain
The full sequence from initial prompt to final usable output in an Agent task. Stability of this chain — no disconnections, no wrong tool picks — determines whether the result is production-ready or requires manual restart.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗