VynarisEarly betaGet your API key

Blog · 2026-08-27 · Vynaris Team

GLM-5.3-Flash costs $1.10 per 1,000 8k/2k calls until September 9

$1.10 per 1,000 calls doubles to $2.20 after GLM-5.3-Flash's promotion ends September 9. Prices verified 2026-08-27.

GLM-5.3-Flash costs $1.10 per 1,000 calls with 8,000 uncached input and 2,000 output tokens. That bill doubles to $2.20 when its 50% launch promotion ends at 2026-09-09 16:00 UTC. The discount is useful; the deadline is the routing decision. Prices verified 2026-08-27.

TL;DR

Verdict table

Decision receipt                                          Promotional rate  After 2026-09-09 16:00 UTC  Change
--------------------------------------------------------  ----------------  --------------------------  ------
Input per 1M tokens                                       $0.075            $0.15                       2x
Cached input per 1M tokens                                $0.015            $0.03                       2x
Output per 1M tokens                                      $0.25             $0.50                       2x
1,000 calls, 8k uncached input + 2k output each           $1.10             $2.20                       +$1.10
1,000 calls, all 8k input tokens cached + 2k output each  $0.62             $1.24                       +$0.62

Buyers should benchmark now, then budget at the list rate. A temporary sticker is a poor permanent forecast.

The $1.10 bill, line by line

Z.AI's live pricing page lists the promotion and its end time. It states 24:00 on September 9 in Singapore time. That equals 2026-09-09 16:00 UTC.

For 1,000 uncached calls, input usage is 8 million tokens. Output usage is 2 million. The 8k input is a disclosed workload shape, not a claim about the model's context window. The promotional bill is:

8M × $0.075/M + 2M × $0.25/M = $0.60 + $0.50 = $1.10

After the promotion, the same workload becomes:

8M × $0.15/M + 2M × $0.50/M = $1.20 + $1.00 = $2.20

That is $0.0011 per call now and $0.0022 later. The absolute increase is $1.10 per 1,000 calls. At one million calls, it is $1,100.

The cached case assumes every input token qualifies for the listed prompt caching rate. It is a boundary, not a forecast. Real cache hits depend on stable prefixes, request order and provider rules. Use the calculator with your own input token cost, output length and call volume.

The quality receipts earn a test, not a migration

The launch report pairs the low rate with public benchmark results. Z.AI reports an Artificial Analysis Intelligence Index v4.1.1 score of 57 at $0.045 per task under the discount.

It also reports 63.4 versus GLM-5.2's 46.2 on DeepSWE v1.1. That is a 17.2-point gap. On AutomationBench, it reports 48.8 versus 26.2, a 22.6-point gap.

Those scores address a real objection: the cheaper model is not presented as price-only. They still do not price our disclosed 8k/2k shape. The $0.045 benchmark task uses its own token mix. We should not reverse-engineer that workload from one total.

The provider published these comparisons. The Artificial Analysis result is third-party, while several other claims appear in the provider's launch report. Neither predicts your acceptance rate, latency or tool-call accuracy.

Our GLM-5.3 quality-adjusted cost analysis shows the safer pattern. Divide spend by passed tasks on your own eval. Our earlier GLM-5.2 model comparison supplies a broader sticker-price baseline.

What should change before September 9

First, tag GLM-5.3-Flash spend separately. A blended provider invoice can hide the step-up.

Second, run a representative sample before the deadline. Measure cost per task, accepted outcomes and tail latency. A leaderboard score cannot set your production threshold.

Third, set two budget alerts. The first uses today's $0.075/$0.25 rates. The second uses $0.15/$0.50. If the second alert breaks the unit economics, the workload needs a smaller output, a higher cache-hit rate or a different route.

Finally, keep the promotion end in UTC in code and procurement notes. “September 9” without a timezone is how an avoidable invoice becomes a postmortem.

Honest tradeoff: the cheapest route may be operationally expensive

Do not route models on this promotion alone. Vendor approval, data location and incident response can outweigh $1.10 per 1,000 calls. A failed coding-agent task can also consume reviewer time that dwarfs the API bill.

The opposite mistake is ignoring a time-boxed test window. The promotion halves a controlled evaluation bill. Use it to collect evidence. Then make the production decision against list rates and a quality-adjusted denominator.

Sources