If your code points at kimi-k3 today, the K3.1 Pro migration is one string. The base URL, the key, the request shape and the billing model stay the same. What changes is the model ID, the price, and on Moonshot's own benchmark, the score.
| K3 | K3.1 Pro | |
|---|---|---|
| Model ID | kimi-k3 |
kimi-k3.1-pro |
| Context | 1,048,576 tokens | 1,048,576 tokens |
| Official input / output / cached | $3.00 / $15.00 / $0.30 | $2.00 / $10.00 / $0.20 |
| Wallaby input / output / cached | see the pricing page | $1.80 / $9.00 / $0.18 |
| Kimi Code Bench 3.0, long-horizon | 55.95% | 61.2% |
| Visual input | yes | yes |
Wallaby's K3.1 Pro line is the official list price with 10% off, and the endpoint opens shortly. The K3 numbers are Moonshot's list prices; our K3 rate card is on the pricing page.
What actually changes in code
One line:
model = "kimi-k3.1-pro"
Everything else in an OpenAI-compatible client stays. Three cases need a second look:
- If you pin model names in config files, search for
kimi-k3and decide per file. - If a gateway routes by model name, add the new ID before you move traffic.
- If you log model names into billing or evaluation tables, add a column rather than overwriting the old one. You will want the K3 rows for comparison.
What the price difference looks like
Take one agent turn: 50,000 input tokens, 4,000 output tokens, 70% cache hits. At official rates, K3 costs $0.1155 and K3.1 Pro costs $0.0770. At our rate card, the same turn is $0.0693.
Output carries the gap. A model that produces 4,000 output tokens costs $0.040 on K3.1 Pro and $0.060 on K3 at output rates alone, before any input tokens are counted.
What we measured, and what we did not
We ran K3.1 Pro on a narrow skill-following test in October. With no skill file it passed 4 of 10 clear tasks; with a skill file it passed 9 of 10. That test measures instruction scaffolding, not general capability, and we published it with that boundary.
We did not run K3 on the same 5,315 multilingual tasks. The multilingual result is a one-arm result for K3.1 Pro. We are not going to turn it into a head-to-head.
When to switch now, and when to wait
Switch now if you need visual input with a 1M context, if your workload is output-heavy, or if you are starting a new integration with no K3-specific tuning.
Wait if your evaluation suite passes on K3 and the volume is small. The per-turn difference above is 4.6 cents. At 10,000 turns a month that is about $460; at 100 turns it is $4.60.
Migration checklist
- Add
kimi-k3.1-proto your allowlist. - Send one prompt you already have a K3 answer for, and compare output length before you compare quality.
- Check the output-token budget in your client. Reasoning effort changes how many tokens the model spends before it answers.
- Keep the K3 model ID in your config for a week and route by config, not by code edit.
- Re-run 20 cases from your evaluation set before you move real traffic.
What does not change
Billing stays per request with itemized statements. The endpoint stays OpenAI-compatible. The model ID is the only new string.
Our K3.1 Pro endpoint opens shortly at $1.80 input, $9.00 output and $0.18 cached input per million tokens, 10% below the official list. The multilingual test and the rate card are up now.