My cheap coding plan ran out in 20 days, and cost more per merged change than GPT-6.1 Sol

Short version: I paid ¥329 (about $46) a month for Xiaomi's MiMo Token Plan Pro and burned through it in 20 days. Per accepted change, MiMo came out at about $0.81 and GPT-6.1 Sol at $0.71. The price per change was close. The real gap was rework: Sol passed review on the first try 71% of the time, MiMo 16%.

The setup

I run a small dispatcher called agent-dispatch. Claude plans the work, writes the orders and reviews the results. A cheaper model does the coding. Each order runs in a macOS sandbox on its own git worktree, then the dispatcher runs tests and lint, then Claude Sonnet reviews the diff. I merge only what passes.

On October 14 I bought MiMo's Pro plan: ¥329 a month for 38 billion credits. The idea was simple. Chinese models are cheap, so let one do the bulk coding and save my Claude quota for planning and review. Claude quota is the most expensive thing in my stack.

What happened

The plan ran from October 14 to November 13. It was empty by November 3.

MiMo Code keeps a local SQLite database. Applying Xiaomi's published credit rates to it, those 20 days come to about 42.7 billion credits, 93% of it from the dispatcher. (Usage between midnight and 8am Beijing time is billed at 0.8x, which is why the plan lasted until November 3.)

In those 20 days MiMo took 87 real orders and ran 238 attempts. 53 orders were accepted. 14 passed on the first try.

When the credits ran out I switched the default lane to GPT-6.1 Sol at medium reasoning, paid per token through the API. From November 3 to 12, Sol took 31 orders and ran 44 attempts. 25 were accepted, 22 on the first try. The API bill was $17.82.

Cost per accepted change

MiMo: 93% of ¥329 is ¥306, about $43. Divided by 53 accepted orders, that's about $0.81 each, at 4.5 attempts per accepted order.

Sol: $17.82 / 25 = $0.71 each, at 1.8 attempts.

The model bill alone is close. Rework is where it diverges. Every attempt runs the test suite, and most also get a Claude review. I sampled one: a Sonnet review cost about $0.20, paid from my Claude subscription. MiMo's extra 2.7 attempts per accepted order add roughly $0.54 in review alone. And time: Artificial Analysis lists MiMo-V2.6-Pro at 35.6 minutes per task, Sol (medium) at 4.2.

Surprise 1: the plan is barely a discount

I assumed the monthly plan was heavily subsidised. It isn't. MiMo-V2.6-Pro pay-as-you-go is ¥3 per million uncached input tokens and ¥6 per million output tokens. The plan's credit rates work out to exactly 100 million credits per yuan, so 38 billion credits is ¥380 of pay-as-you-go usage.

¥329 for ¥380 of tokens is 13% off. Unused credits expire. Run out mid-month and it doesn't roll over to pay-as-you-go. You just wait. Compared with Claude it's still very cheap. Compared with its own API price, it's a prepaid card.

Surprise 2: my own ledger undercounted by 2.6x

The dispatcher records token usage for every order. When I first totalled the month from my own ledger, MiMo had used about 15 billion credits, less than half the plan. I nearly concluded I'd bought too big a plan.

Only after the credits actually ran out did I check MiMo Code's database. The dispatcher's share for the same period was 39.7 billion. My ledger was off by about 2.6x. The main cause: it only counted usage events from MiMo's output stream, which covers the main session. Subagents write their usage to child sessions in the database, and those never appear in the stream. The fix is in progress.

If you log usage yourself, reconcile it against the provider's numbers. Otherwise every cost analysis you build on it is wrong.

The leaderboard agrees

Artificial Analysis model leaderboard, November 13, 2026 (Intelligence Index / cost per task / time per task):

At the same $0.26 per task, Terra (high) scores 12 points higher than MiMo and finishes 8x faster. The only Chinese model at Sol-medium level is Kimi K3.1 Pro, at 10x the cost and 9x the time. This is a general intelligence index, not a coding benchmark, but it points the same way as my 16% vs 71%.

What I run now

Sol (medium) is the default. Rework and hard orders move up to Sol (xhigh). The few dozen yuan left in my MiMo account now sit on pay-as-you-go as a fallback for when OpenAI is unreachable. No more monthly plan.

I also tried GPT-6.1 Terra (xhigh) on four replayed orders. It used 3 to 4x the tokens Sol (medium) did, so half the token price saved nothing. At my current volume I expect $60 to $100 a month.

Caveats: the samples are small (87 orders vs 31), the two periods had different work, and MiMo's API was stalling for part of late October, which hurt its pass rate. Treat my numbers as an example of the method, not a verdict.

What I'd suggest

1. Compare models by cost per accepted change, including retries, review cost and wall-clock time, not by price per token.
2. Before buying a token plan, convert its quota into pay-as-you-go money and compare that with what you actually use.
3. Reconcile your own usage logs with the provider's at least once a month.

I also do this for other teams as a fixed-price routing audit: cost per accepted change across your models, a reconciliation of your usage logs against provider data, and a routing recommendation. It's US$1,500. Reach me at info@zhuojun.li.

Happy to compare notes if you're running something similar.