GLM-5.3-Flash after the promo: $0.15/$0.50, MIT, and what Z.AI's numbers say about inference cost

Z.AI's launch promo on GLM-5.3-Flash ends today, September 9, at 24:00 UTC+8. From tomorrow the list price is $0.15 per million input tokens and $0.50 per million output, with cached input at $0.03. During the promo Artificial Analysis measured the model at roughly $0.045 per task for an Intelligence Index score of 57. Anyone who built unit economics on that number gets a 2x correction overnight, and that is the least interesting thing about this model.

What shipped on August 26

GLM-5.3-Flash is a natively multimodal mixture-of-experts model, 320B total parameters with 18B active, a 1M-token context, an MIT license and weights on Hugging Face. It ran anonymously for a week as Ox Alpha on OpenRouter and OpenCode before Z.AI claimed it, and it was served on domestic Chinese chips. Z.AI says it beats GLM-5.2 across its benchmarks at about a tenth of the price and lands within 0.5 points of Claude Opus 4.8 on the company's internal coding benchmark. This is the Flash variant of GLM-5.3, whose full weights were held back on August 20 over cyber-safety concerns.

The active parameter count is the number to hold onto. 18B active means decode cost per token is closer to a mid-size dense model than to a 320B one, which is how a vendor can put a near-Opus coding claim next to a fifty-cent output price and not be lying about either.

What Z.AI's own numbers say about inference cost

The pricing is not a loss-leader stunt, or at least the company's financials suggest it does not need to be. Z.AI reported H1 revenue of about 954 million CNY, roughly $142 million, up around 400% year over year. The cloud and API business was about 825 million CNY, up 27x. The company says it runs large clusters on domestic Chinese silicon and that its inference cost is falling.

Put those together: a lab with an API-first revenue line, growing at that rate, that does not buy Nvidia at Nvidia prices, and that ships an 18B-active model. The $0.15 / $0.50 list price is a consequence of that cost structure, not a marketing budget. It is also a floor that Western API providers cannot match on the same hardware economics, which is why this model matters for routing even if you never use it.

A promo tells you what a vendor wants you to try. The list price tells you what their inference actually costs. Model your pipeline on the second number.

Calendar every promo you depend on

The operational lesson is boring and I keep watching teams miss it.

• Write down the expiry. Every discounted model in your router gets a calendar entry a week before the promo ends and a re-forecast attached to it. GLM-5.3-Flash is a 2x step today; other launch discounts this quarter have been steeper.
• Model on list price, spend at promo price. Treat the discount as margin, not as the plan. If the tier only works at $0.045 per task, it does not work.
• Keep the fallback warm. At $0.15 / $0.50 the model is still cheap; the question is whether it is still the cheapest that reliably clears your quality bar. Compare against the current price table and against GLM-5.1 if you already have it in production.
• Use the cache price. $0.03 cached input is a fifth of the list rate. For agent loops with a stable system prompt, that is where most of the savings survive the promo.
• Log provenance. Z.AI was named in the September 8 NSA/CISA/FBI distillation advisory. Nothing in it bans the model, but a regulated client will ask, and cheap is not a governance answer.

The honest gap

The benchmark claims are Z.AI's own. The 0.5-point gap to Opus 4.8 is on an internal coding benchmark I cannot inspect, and the claim of beating GLM-5.2 across the board has the same provenance problem GLM-5.2's numbers had. The Artificial Analysis figure is independent but is an aggregate index, not your workload. The financials come from Caixin's reporting of company figures, not audited statements. What is verifiable: 320B with 18B active, 1M context, MIT, weights on Hugging Face, and a list price you can read off the site tomorrow morning. Build your eval on those and let the vendor numbers be a hypothesis.