GLM-5.3-Flash costs $0.075 per million input tokens and $0.25 per million output tokens during a promotion that runs until September 9, 2026, after which list pricing of $0.15 and $0.50 applies. The free Ox Alpha preview ended on August 26, 2026 when the model was revealed. This page has the full rate card, worked cost examples, and an honest comparison with what you were paying before.
Rate card
| Per 1M tokens | Promotional (until Sept 9, 2026) | List price (from Sept 10) |
|---|---|---|
| Input | $0.075 | $0.15 |
| Cached input | $0.015 | $0.03 |
| Output | $0.25 | $0.50 |
| Reasoning tokens | Billed as output. Reasoning is always on for this model. | |
| Image / video input | Billed as input tokens | |
| Minimum spend / subscription | None — pay per token | |
The promotion ends at 24:00 UTC+8 on September 9, 2026. Prices exactly double at that point; nothing else about the model changes. The live figures the chat on this site uses are read directly from the OpenRouter listing, so if anything moves before then, the spec strip on the home page will show it before this page does.
The cached-input discount is the important number
Cached input is billed at one fifth of the normal input rate — $0.015 per million during the promotion. On a model with a million-token context window, that is not a rounding detail, it is the difference between a viable workflow and an expensive one.
The pattern that benefits: send a large, stable prefix (a codebase, a specification, a document set) and vary only the question at the end. The prefix is cached after the first call and the rest of the session bills at the cached rate. The pattern that does not benefit: rebuilding the prompt from scratch each turn, or putting the variable part at the beginning. If you are designing an agent around this model, arrange the context so the unchanging material comes first.
Worked examples
All figures at promotional pricing; double them for list pricing from September 10.
| Task | Tokens | Cost |
|---|---|---|
| A single question about a mid-sized codebase | 200,000 in · 8,000 out | $0.017 |
| Filling the entire context window once | 1,048,576 in · 32,000 out | $0.087 |
| Long report from a large document set | 500,000 in · 100,000 out | $0.063 |
| Ten follow-up questions on a cached 200k prefix | 200,000 in, then 9 × cached · 5,000 out each | $0.054 |
| A month of steady agent work | 50M in · 5M out | $5.00 |
The last row is the one worth sitting with. Fifty million input tokens a month is a substantial agentic workload — continuous code review, document processing, a coding assistant running all day — and it costs about the same as a sandwich. At list pricing after September 9 it costs about two sandwiches.
What changed when the preview ended
- You now need credit. During the preview a $0 balance worked. It does not any more — an empty account returns a 402 error.
- The old model ID is gone.
stealth/ox-alphais delisted with no redirect. Anything still calling it fails outright rather than silently rerouting to the paid model, which is inconvenient but much better than a surprise bill. - You know who you are paying. Z.ai, under standard commercial terms rather than the Stealth Program’s data-for-access arrangement.
- You can opt out entirely. The weights are MIT-licensed on Hugging Face, so self-hosting is a real alternative to paying per token.
What each provider charges
Because the weights are MIT-licensed, anyone can host GLM-5.3-Flash — and within about a day of the release, a dozen providers were doing exactly that. OpenRouter routes across all of them, so the price you actually pay depends on which one serves your request, and if a provider errors OpenRouter fails over to another healthy one rather than failing your call.
| Provider | Input / 1M | Output / 1M | Cached input / 1M | Status |
|---|---|---|---|---|
| GMICloud | $0.075 | $0.25 | $0.015 | Degraded |
| Novita | $0.075 | $0.25 | $0.015 | Serving |
| Z.AI | $0.075 | $0.25 | $0.015 | Serving |
| Modal | $0.15 | $0.5 | $0.03 | Degraded |
| BaseTen | $0.15 | $0.5 | $0.03 | Serving |
| Cloudflare | $0.15 | $0.5 | $0.03 | Serving |
| DeepInfra | $0.15 | $0.5 | $0.03 | Degraded |
| Io Net | $0.15 | $0.5 | $0.03 | Degraded |
| Parasail | $0.15 | $0.5 | — | Serving |
| Reka | $0.15 | $0.5 | $0.03 | Serving |
| Together | $0.15 | $0.5 | $0.03 | Serving |
| Venice | $0.15 | $0.5 | $0.03 | Degraded |
Two things are worth noticing in that table. The first is the shape of the spread: a small group sits at the promotional rate while most independent hosts sit at roughly double it. The second is what that implies about September 10. Z.ai’s post-promotional list price is $0.15 input and $0.50 output — which is exactly where the independent providers already are. The promotion is not a discount off some future premium; it is a temporary position below what the open market has settled on. When it ends, Z.ai stops undercutting the field and joins it.
Practically, that means the promotion’s end is less dramatic than “prices double” makes it sound. You are not losing access to a cheap model — you are losing first pick of the cheapest host. Routing will still find you the floor of whatever the market offers. The floor itself is what moves.
A caveat on reading the table: cheapest is not automatically best. Providers differ in the maximum output they will serve, in throughput, and in whether they are currently healthy at all. A row marked Degraded is not taking traffic right now, which is precisely the situation failover exists to handle.
How it compares
Against the flagship models from the major commercial families — ChatGPT, Claude, Gemini and Grok — GLM-5.3-Flash sits at the very bottom of the price range while offering a million-token context window, native multimodal input and open weights. Those flagships are generally stronger on the hardest reasoning tasks, and none of this makes GLM-5.3-Flash the right default for every workload. It makes it very hard to beat on cost per token for high-volume, long-context work. Our comparison guide goes family by family, and there is a dedicated page for how it relates to the rest of the GLM line.
One in-family comparison is worth making explicitly: z-ai/glm-5.3, the larger text-only sibling, is substantially more expensive per token. If your workload is long-context or multimodal rather than maximally difficult reasoning, Flash is usually the better economics.
If your volume is very large, or your data cannot leave your own infrastructure, the MIT-licensed weights make self-hosting a real option — see self-hosting GLM-5.3-Flash for hardware requirements and a breakeven analysis against the prices above.
Free alternatives
If you want zero cost rather than low cost: OpenRouter maintains a rotating set of free models, and the chat on this site lets you select one from the model switcher without any credit on your account. You can also run GLM-5.3-Flash yourself under its MIT licence — free in licensing terms, though you pay for the hardware. Both routes are covered in the setup guide, and the preview timeline page has the full story of what ended and when.
Frequently asked
How much does GLM-5.3-Flash cost?
$0.075 per million input tokens and $0.25 per million output tokens until September 9, 2026, then $0.15 and $0.50. Cached input is $0.015 per million during the promotion.
Is Ox Alpha still free?
No. The free preview ended on August 26, 2026 when the model was revealed as GLM-5.3-Flash.
Do I need to add a credit card?
You need credit on your OpenRouter account, which does require a payment method. A 402 “insufficient credits” error means your balance is empty.
When does the promotional pricing end?
September 9, 2026 at 24:00 UTC+8. List pricing of $0.15 input and $0.50 output applies from then.
Is there a usage limit?
No published cap beyond your credit balance. Heavy use can return 429 rate-limit responses; retry with backoff.
Sources and verification
Last substantively reviewed: August 27, 2026. Reviewed by: OxAlpha.chat Editorial Team.
How we verified this page
Prices were copied from Z.ai and compared with OpenRouter’s live provider table on August 27, 2026. Worked examples are our calculations using the displayed per-token rates. Prices and provider availability can change after publication.
Primary sources
- Official pricing: Z.ai API pricing — states the standard rate, cached-input rate and temporary promotional period.
- Live platform pricing: OpenRouter GLM-5.3-Flash listing — shows current provider prices and route availability.
- Endpoint metadata: OpenRouter endpoint-list reference — defines the provider-level price, context and status fields used by live tables.
- Routing rules: OpenRouter provider selection documentation — explains ordering, fallbacks and route controls.
- License: Official MIT license file — supports the self-hosting alternative discussed on this page.
Found a factual error or a changed price? Send us the page URL and supporting source; corrections follow our editorial policy.