On Kimi K3.1 Pro, thinking tokens are billed as output. That is the whole cost story of the effort dial in one sentence: at $10 per million output tokens on official rates ($9 on Wallaby), a run that thinks for 30,000 tokens pays $0.30 before it writes a single visible word.
Disclosure: Wallaby Token sells K3.1 Pro. The rates are Moonshot's published list; the arithmetic below is ours.
The ladder, priced
Take a code-review task that writes 1,200 visible tokens and reads 40,000. Only the thinking budget changes:
| Thinking budget | Total output billed | Output cost, official | Output cost, Wallaby |
|---|---|---|---|
| 2,000 | 3,200 | $0.0320 | $0.0288 |
| 8,000 | 9,200 | $0.0920 | $0.0828 |
| 24,000 | 25,200 | $0.2520 | $0.2268 |
| 48,000 | 49,200 | $0.4920 | $0.4428 |
Input stays constant at $0.08 for the 40,000 tokens. So going from the 2,000-token budget to the 48,000-token budget multiplies the total bill by roughly 4.4×, and 97% of the difference is thinking, not the answer.
What we measured on K3
In October we ran the same review task four times at four effort tiers through our own gateway and kept the itemised bills. The tiers produced 9,644, 17,094, 23,806 and 35,946 output tokens, and the charges ran $0.42, $0.72, $0.86 and $1.43. The tier labelled "None", the one people pick expecting savings, produced the most tokens and the biggest bill, 3.4× the cheapest run. On a hard prompt, "no budget" is a soft switch, not a cap: the model reasons anyway, and burns tokens on trial and error that a budget would have cut short.
The full bills, screenshots and method are in the four-budget post. Those runs were on K3 at K3 rates; the K3.1 Pro arithmetic above is the same ladder at the new rate card, not a fresh benchmark.
Reading your own numbers
Every OpenAI-compatible response can carry a usage block. On our gateway the reasoning count sits next to the completion count, so one log line answers the question: how many of the output tokens were thinking?
- Thinking share = reasoning tokens / output tokens. Above 80% is normal for reasoning models on hard tasks; it is not waste by itself.
- Cost per accepted answer. Divide the bill by the number of answers you kept. This is the number that decides whether the dial is set right, because a cheaper run that produces two rejected answers is not cheaper.
- Output:input ratio. On K3.1 Pro output is five times the input rate. A task that reads 200,000 tokens and thinks 40,000 is an input problem; a task that reads 4,000 and thinks 40,000 is an output problem, and the dial is the only lever.
Available now
- Model IDs:
kimi-k3.1-proandkimi-k3.1-pro-highspeed, served athttps://api.wallabytoken.com/v1behind an OpenAI-compatible chat-completions endpoint. - Context: 1,048,576 tokens, with reasoning output billed as output.
- Cache: repeated prefixes bill at $0.18 per million on both variants.
- Price: $1.80 / $9.00 per million input / output tokens on the standard variant, $3.60 / $18.00 on Highspeed — a flat 10% below Moonshot's list, standing rate with no promo code and no end date.
- Start: $0.50 trial credit on sign-up with no card, then top-ups by card, Apple Pay or Google Pay.
The rule
Set the budget by task class, not by mood: cheap tier for extraction, classification and edits you will review anyway; higher tiers for the tasks where a wrong answer costs more than the tokens. Then check the bill once a week: the tier names are the same across providers, the token counts are not.
The rate card behind these numbers is in the K3.1 Pro pricing guide; the cache side of the same bill is in the cache pricing post.