Short answer: OpenRouter listed GLM-5.2 at $0.007 per million input tokens and $6.80 per million output tokens on November 13, 2026, with a $0.006 cache-read rate and a 1,048,576-token context. Input is roughly 971x cheaper than output, so a 10M-input, 1M-output month costs about $6.87, while a 1M-input, 10M-output month costs about $68.01 for the same total token count.
| Route | Prompt | Completion | Note and meter |
|---|---|---|---|
| z-ai/glm-5.21,048,576 context | $0.007 | $6.80 | The standard route in this retrieval. |
| z-ai/glm-5.2:batchbatch variant | $0.298 | $0.916 | Batch input is 42x the standard input rate at this retrieval. |
| glm-5-2 on fireworks-aiserverless-standard | $1.40 | $4.40 | APIDir accepted observation checked 2026-08-22. |
| glm-5-2 on together-aiserverless | $1.40 | $4.40 | APIDir accepted observation checked 2026-08-22. |
Reading the ratio before the rate
At $0.007 input and $6.80 output per million tokens, GLM-5.2's ratio is about 971:1. For comparison, the same retrieval puts Qwen3.8 Max at 3:1, Claude Sonnet 5 at 5:1, GPT-5.4 at 6:1, and Kimi K3 at 83:1. A ratio that extreme means the workload shape, not the model quality, determines whether the route is cheap.
Two months with the same total token count can differ by an order of magnitude. Ten million input tokens and one million output tokens cost $0.07 + $6.80 = $6.87. One million input and ten million output cost $0.007 + $68.00 = $68.01. Same volume, 10x the bill. Any comparison that reports a single blended rate per million for this model is hiding the only number that matters.
Where the route is genuinely strong
Input-heavy work is the fit: classification over long documents, retrieval-augmented extraction with short structured answers, log and transcript triage, and batch enrichment that returns a field or a label. In those workflows output per request is small and predictable, so the $6.80 meter contributes little and the cheap input meter carries the volume.
The one-million-token context supports long inputs, but context is a ceiling rather than a budget target. Retrieval precision, latency, and the provider's rate limits still bound the workload. Test on documents from your own corpus rather than a generic benchmark, because a cheap route that fails your extraction task is not cheap once human review is added.
- Fit: classification, extraction, tagging, log triage, and enrichment with short structured answers.
- Fit: batch review of long documents where the answer is a field, label, or short JSON object.
- Poor fit: chat, copywriting, code generation, and anything whose output length is set by the user.
- Verify the route's latency and rate limits on your own volume before assuming the input price sets the budget.
Where it loses: long-form output
Any product that generates prose, code, or long structured payloads pays the $6.80 output rate on every token. A 1,200-token answer costs about $0.008; at 500,000 answers a month that is $4,080, which is more than the same volume would cost on Qwen3.8 Max at $6.00 output and far more than a route with a lower completion rate.
The decision rule is simple: estimate average output tokens per accepted response and multiply by monthly volume. If that product exceeds the alternative route's total, GLM-5.2 is the wrong choice regardless of how attractive the input price looks. Cap output length in the product contract and measure it from provider usage fields, not from a UI character count.
Comparing GLM-5.2 with another route therefore needs one shared assumption: the output length your product actually ships. Fix that number, run both routes at the rates you retrieved, and compare totals. Averaging the two routes' input and output prices into a single blended rate destroys exactly the information that makes the comparison useful, because the ratio, not the average, is what separates a reading workload from a writing one.
The cross-provider gap is large
APIDir's accepted observations for GLM-5.2 on Together AI and Fireworks AI are $1.40 input and $4.40 output, checked August 22, 2026. The OpenRouter route retrieved in November lists $0.007 input and $6.80 output. Those are not the same product priced differently; they are different routes whose rates moved, and the input gap is two orders of magnitude.
That is a freshness warning, not a recommendation. A rate retrieved today can be a promotional or restructured price, and the older accepted row may describe a configuration that no longer exists. Recheck the route you will actually call, capture the response with a timestamp, and reconcile a small invoice before scaling. Do not average the two rows into a blended assumption.
A practical evaluation plan
Run 200 representative tasks through the candidate route, record input and completion tokens per task, the pass rate against your rubric, and the latency distribution. Compute cost per accepted task using the retrieved rates plus any human review time the failures require. Then compare that number with the current route on the same task set. Cost per accepted task is the only figure that survives contact with production.
Keep the route behind an adapter so the model identifier, base URL, and any provider-specific parameters can change without touching product code. Record the retrieved price date in the configuration comment or a runbook. When the next catalog change lands, the adapter makes a re-test cheap and the budget review factual.
See how output changes the GLM-5.2 bill
Use the retrieved route rates, then edit the output length to watch the bill move. The input meter barely registers at these prices.
Planning result
Review the example inputs, replace each rate with the exact current value from the linked source, then calculate. APIDir does not send these values to a provider.
Questions developers ask
How much does GLM-5.2 cost per million tokens?
OpenRouter listed $0.007 input and $6.80 output per million tokens on November 13, 2026, with a $0.006 cache-read rate. Recheck the route before relying on either number.
Is GLM-5.2 good for long-form generation?
The output rate is the highest in APIDir's current text-route comparison, so long-form products should compare cost per accepted answer with a cheaper-completion route before choosing it.
Why is GLM-5.2's input price so low?
The retrieved catalog lists it that way. Treat a rate that far from other routes as an observation to verify with a live call and an invoice sample, not as a permanent market fact.
Does the batch route save money?
It depends on the workload. At this retrieval the batch row lists higher input and lower output than the standard route, so input-heavy jobs may pay more.
Sources and check dates
Every rate in this guide was retrieved from the linked page on Nov 13, 2026. Prices can change; recheck the source before purchasing.