benchr
benchr rechecked Moonshot's pricing page and the official model card on September 8, 2026. Two things in this release are worth a decision, and neither of them is a leaderboard position.
The thinking-token claim is the cost claim
Moonshot says K2.7-Code reduces thinking-token usage by approximately 30%
against K2.6 while improving on long-horizon coding work. Thinking tokens bill as output, and output here is $4.00 against $0.95 input — a ratio of 4.2 to 1.
So a 30% reduction in thinking tokens is not a footnote about efficiency. On a workload dominated by reasoning, it is close to a 30% cut in the expensive half of the bill, and it arrives without a rate change. That is the opposite of the usual pattern, where a newer model costs more per token and you argue about whether it earns the difference.
It is also the claim you can check yourself most easily. Run the same repository task on both, count output tokens, and the ratio is either there or it is not.
High-speed costs exactly double, everywhere
The hosted high-speed variant, Kimi K2.7-Code-Highspeed, called with kimi-k2.7-code-highspeed, is not a different model. Same 262,144-token context, same 1T/32B architecture, and every published rate is precisely twice the base.
| Rate | kimi-k2.7-code | high-speed | Multiple |
|---|---|---|---|
| Input, cache miss | $0.95 | $1.90 | 2.0x |
| Input, cache hit | $0.19 | $0.38 | 2.0x |
| Output | $4.00 | $8.00 | 2.0x |
| Context window | 262,144 | 262,144 | same |
A flat 2x across every line makes the decision unusually clean: you are buying latency and nothing else, at a fixed and legible price. There is no capability argument to have, no context trade, and no benchmark difference to weigh. Either the wall-clock time is worth double, for this workload, or it is not.
Moonshot publishes no maximum-output figure and no license for the high-speed variant on the pricing page, so benchr records neither. The base model's modified-MIT weights are not evidence about a hosted endpoint's terms.
Cache hits are where the input bill actually lives
Cache-hit input is $0.19, exactly a fifth of the $0.95 miss rate. For coding work that repeatedly sends the same repository context, the difference between a well-structured prompt prefix and a poorly structured one is a 5x swing on the input side.
That is a bigger lever than the choice between the two variants for most teams, and it is entirely under your control. Measure the hit rate before comparing models; a low hit rate makes every model look expensive and hides which one is actually cheaper for the work.
The published scores, and what they are
Six figures come with the model card. All six are Moonshot's own; benchr has reproduced none of them.
| Benchmark | Score | K2.6 |
|---|---|---|
| MCP Mark Verified | 81.1 | not published |
| MCP Atlas | 76.0 | not published |
| Kimi Code Bench v2 | 62.0 | 50.9 |
| Program Bench | 53.6 | not published |
| Kimi Claw 24/7 Bench | 46.9 | 42.9 |
| MLS Bench Lite | 35.1 | not published |
Two of the six carry a K2.6 comparison, and those are the informative ones: Kimi Code Bench v2 rises 50.9 to 62.0, Kimi Claw 24/7 rises 42.9 to 46.9. Note that the largest published gain is on Moonshot's own named benchmark, which is the normal shape of a vendor release and a reason to weight your own tasks over any of these.
Who this fits
| If you are | The record says |
|---|---|
| Running long-horizon coding agents and watching output cost | Test it. The thinking-token reduction is the claim, and output is the expensive half |
| Blocked on latency rather than cost | The high-speed variant is a clean 2x for speed alone, with no other difference |
| Required to self-host or audit weights | The base model publishes modified-MIT weights at 1T total, 32B active |
| Sending large repeated repository context | Fix the cache hit rate first; $0.19 against $0.95 is a bigger lever than the model choice |
| Choosing on published benchmarks alone | Do not. Four of six have no prior-generation comparison and the largest gain is on the vendor's own suite |