The price cut is real. The workload has to be input-heavy for it to matter.

MoonshotAI cut Kimi K2.5 input pricing from $0.60 to $0.45 per million tokens, and Z.ai cut GLM 5.1 from $1.26 to $0.97. That is a real production change, but it only changes a buying decision when the application is dominated by input tokens rather than output, latency, or evaluation risk.

The clean version of the saving is simple. At one billion input tokens a month, Kimi K2.5’s cut is worth $150: the old bill was $600 and the new one is $450, using the OpenRouter catalogue prices reported in today’s change scan. At the same volume, GLM 5.1’s cut is worth $290: $1,260 before, $970 after. At ten billion input tokens, those become $1,500 and $2,900 a month. At one hundred billion, they become $15,000 and $29,000. The percentage headline is large; the dollar impact depends entirely on whether your workload already lives at that scale.

This is why price cuts are easy to overread. A 25% input cut on Kimi K2.5 sounds like a platform event. For a retrieval-heavy support system sending long transcripts, policy text and account history into the prompt, it may be one. For a coding assistant, analyst agent or customer-facing chat product whose cost is materially output-side, it may be a footnote. The input line got cheaper; the whole request did not necessarily get cheaper.

The second cut is larger in dollars because GLM 5.1 started from a higher input price. OpenRouter’s measured change was $1.26 to $0.97 per million input tokens, a 23% reduction. On the same billion-token prompt budget, that saves almost twice as much as Kimi K2.5’s cut. That does not make GLM 5.1 the better deployment choice. It says the old price left more room to move.

For teams running production inference, the useful question is not “which model got the biggest percentage cut?” It is “which line of my bill is sensitive to this cut?” If the application prepends a large system prompt, injects retrieval passages, includes prior turns and asks for short structured answers, input price can dominate. If the application generates long reports, code, chain-of-thought-free but lengthy answers, or multi-step agent output, output tokens and wall-clock time can be the real bill.

The catalogue change also landed in a market where measurement is still thin. OpenRouter supplies the listings and prices; Artificial Analysis supplies independent quality, time-to-first-token and throughput measurements. The current catalogue has 396 listed models, but only 151 have independent quality scores and only 57 have a published time-to-first-token measurement. That means 339 listings have no published TTFT at all. They are not instant; they are unmeasured.

That matters because a price cut can make an unmeasured deployment look cheaper than it is. A batch summarisation job may tolerate slow starts. A person waiting in a chat window may not. Artificial Analysis’ measured TTFT distribution runs from 0.56 seconds at the tenth percentile to 36.6 seconds at the ninetieth percentile, with a 1.58 second median. That is not a rounding error around the same user experience. It is the difference between interactive and visibly stalled.

Throughput and first-token latency also should not be collapsed into one “speed” idea. Artificial Analysis measures them separately because they answer different production questions. Time to first token is what a waiting user feels. Tokens per second is how quickly a long answer finishes once it has started. A model can be quick to begin and slow to complete, or slow to begin and fast once running. A cheaper input price does not tell you either number.

There is one more trap in treating the cut as a model recommendation: price moved on specific models, while quality-per-dollar leaders may be elsewhere. In today’s measured catalogue, Z.ai GLM 5.3 Flash leads the value ranking among models with an intelligence index of at least 50, at 58 index points and $0.07 per million input tokens. DeepSeek V4 Flash 0423 follows at 52 index points and $0.09 per million input tokens. Those numbers come from the combined OpenRouter price data and Artificial Analysis quality measurements, and they narrow an evaluation shortlist; they do not prove either model fits a production task.

So the action from the cut is narrower than the headline. If you run an input-heavy workload and Kimi K2.5 or GLM 5.1 is already in your evaluation set, update the cost model immediately. At one billion input tokens a month, use $150 saved for Kimi K2.5 and $290 for GLM 5.1 as the first-order number. At ten billion, multiply by ten. Then test the part the price sheet does not answer: task quality, first-token latency, throughput and provider behaviour under your concurrency.

If your workload is not input-heavy, do not let the percentage cut choose the model for you. It is a better price on one side of the request, not a complete inference budget. The teams that benefit most are the ones that already know their token mix well enough to say which side of the meter just moved.

Both projects are on GitHub and PyPI. Install them.