Everyone says Chinese AI is dirt cheap
5 hours ago
- Kimi K3 is not the cheap option; it costs similarly to Western flagships and performs mid-table on cost per test.
- Chinese models dominate as the 'doer' (coding execution), being 19 times cheaper than Western equivalents for the same results.
- Three traps include hidden billing for reasoning tokens, models failing in real tooling despite leaderboard rankings, and unstable routing on cheap aggregators.
- The optimal setup uses a frontier planner (e.g., Grok-4.5 or GPT-5.6 luna) paired with a cheap Chinese doer (e.g., DeepSeek flash).
- Cost should be measured per verified test pass, not per token.