Kimi K3.1 Pro vs Highspeed: What Does the Faster API Cost?
Highspeed lists double Pro’s rate for regular input and output, yet cache hits cost $0.20 on both. How much extra will your actual workload pay for the speed?
- Pro costs $2 per million regular input tokens and $10 per million output tokens.
- Highspeed charges $4 and $20, but cache hits remain $0.20 for both.
- Compare equal reasoning settings and actual usage when pricing the speed premium.
The speed premium depends on your input mix
The official Kimi API platform describes Highspeed as offering the same intelligence and context size as K3.1 Pro at greater speed. Its regular input and output rates double, while cached input costs the same.
| Per million tokens | Pro | Highspeed |
|---|---|---|
| Regular input | $2 | $4 |
| Default five-minute cache write | $2 | $4 |
| Cache-hit input | $0.20 | $0.20 |
| Output | $10 | $20 |
These are USD rates before tax, checked October 26, 2026. The billing documentation distinguishes the pay-as-you-go API from Kimi memberships and Kimi Code subscriptions.
A 50,000-input, 10,000-output request costs $0.20 or $0.40
Assume no cache hits and that all input is billed as regular input or default five-minute writes. Pro’s input costs $0.10 and its output another $0.10. The equivalent Highspeed request costs $0.40.
Now consider a follow-up with 40,000 cached input tokens, 10,000 regular input tokens and the same 10,000 output tokens:
| Follow-up request | Pro | Highspeed |
|---|---|---|
| 40,000 cached tokens | $0.008 | $0.008 |
| 10,000 regular input tokens | $0.020 | $0.040 |
| 10,000 output tokens | $0.100 | $0.200 |
| Total | $0.128 | $0.248 |
The premium is $0.12. This follow-up calculation excludes the earlier cost of creating the cache and assumes no new one-hour writes. It illustrates why the total bill need not double, even though two prominent rates do.
Split input before applying rates
Kimi’s cache guide defines Chat Completions input as three mutually exclusive portions: reads, writes and the uncached remainder. Applying the regular rate to all usage.prompt_tokens and then adding write charges counts some tokens twice.
Read prompt_tokens_details.cached_tokens and prompt_tokens_details.cache_write_tokens to separate those portions. The default write TTL is five minutes; one-hour writes have a separate rate. A cache hit does not incur another write charge.
Hold reasoning effort constant when comparing speed
Kimi’s release announcement claims a 2–3× speed improvement. That does not imply an agent finishes an entire workflow, including external tools and tests, in one-third of the time. Compare completion time and cost on the same requests.
The model guide provides another latency control: reasoning_effort supports low, high and max, defaulting to max. Thinking remains enabled. Kimi suggests lowering effort when reasoning takes too long.
Start with Pro when minimizing cost is the priority. For an interactive task, compare Highspeed at the same effort setting. Test reduced effort separately, checking answer quality and token usage alongside the wait.
An API endpoint is available before self-hosting weights
The official model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Kimi says full weights will be released by November 4, 2026. Hosted access and downloading the model for your own infrastructure have different timelines.
For an API decision today, split your workload into regular input, cache activity and output. Then price the faster option against the waiting time it actually removes.