跪拜 Guibai
← All articles
AI Coding · Flutter

Qwen 3.8 Max Burned a 5-Hour Token Quota on One Flutter Task That a Free Model Finished in 3 Minutes

By 程序员老刘 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Benchmark scores that don't test your actual stack and workflow are a costly proxy. A model can rank near the top on LiveBench coding yet be economically unusable for Flutter work because its verbosity burns through paid quotas before finishing the job.

Summary

A straightforward Flutter integration job — wiring three finished modules into a main workflow — consumed an entire 5-hour token plan on Qwen 3.8 max without completing. The same task finished in 3 minutes and 17 seconds using opencode's free model. The root cause was excessive verbosity: the model's reasoning chains grew so long across rounds that token limits were exhausted before the work was done.

LiveBench scores place Qwen 3.8 max just below Claude and GPT-5.6 in overall and coding capability, but those benchmarks test Python and web tasks, not Flutter. The developer's actual workflow — using a top model to decompose requirements, then a mid-range model to build each feature — doesn't map cleanly to any single benchmark dimension. Cost-effectiveness rankings confirm the coding inefficiency: the model's strength lies in logic and data processing, not programming.

A practical takeaway emerges: the remaining token quota handled typo fixes, formatting, and GitHub publishing for the very article describing the failure, reinforcing that Qwen 3.8 max performs better on non-coding chores. Low-tier paid plans that can't survive a single long conversation hold no practical value when competitors offer generous free tiers.

Takeaways
Qwen 3.8 max consumed an entire 5-hour token plan on a single Flutter module-integration task and still didn't finish.
opencode's free model completed the identical task in 3 minutes and 17 seconds.
The failure mode was excessive verbosity: long reasoning chains carried across conversation rounds rapidly exhausted the token budget.
LiveBench coding benchmarks don't include Flutter, so scores on Python and web tasks mispredict Flutter performance.
LiveBench cost-effectiveness rankings show Qwen 3.8 max is far less efficient at coding than at logic and data processing tasks.
The developer's remaining token quota successfully handled non-coding chores like typo fixes, formatting, and GitHub publishing for the same article.
Low-tier paid plans that can't complete one long conversation are impractical when vendors like Agnes, SenseTime, OpenRouter, and opencode offer substantial free quotas.
Conclusions

Benchmark-to-project mismatch isn't just about language; it's about workflow. A model optimized for single-shot code generation may fail badly in multi-turn, context-accumulating agentic loops.

Verbosity is a hidden cost driver that benchmark scores don't capture. A model can be both capable and economically non-viable if its reasoning style inflates token consumption per task.

The fact that the same model handled text-editing chores efficiently but failed at coding suggests the token-burn problem is specific to how it approaches programming tasks, not a universal trait.

Concepts & terms
LiveBench
A benchmark that uses non-public, periodically refreshed questions to evaluate AI models, reducing the risk of score inflation from training-data contamination.
Agentic coding
A coding workflow where an AI model operates with some autonomy — planning, executing, and iterating across multiple steps — rather than responding to a single prompt.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗