GLM-5.3-Flash is currently half price. That ends on September 9, 2026, and when it does, every cost projection built on today’s rates doubles.
This post covers what the model costs now, what it costs after, how that compares to the alternatives, and how to work out whether the difference actually matters for your workload. All figures were verified on August 27, 2026, and pricing pages change, so confirm against your provider before you commit a budget.
The rate card
| Launch price (through Sept 9, 2026) | List price (after) | |
|---|---|---|
| Input | $0.075 / 1M tokens | $0.15 / 1M tokens |
| Output | $0.25 / 1M tokens | $0.50 / 1M tokens |
| Cached input | $0.015 / 1M tokens | $0.03 / 1M tokens |
Artificial Analysis publishes a blended figure of $0.10 per million tokens using a 7:2:1 cache-hit to input to output ratio. That is a useful single number for comparison shopping, though your own ratio is what determines your bill.
What this costs in practice
Abstract per-million rates are hard to reason about, so here are three concrete workloads at list price, which is the number that matters for planning past September 9.
A support classifier. 100,000 tickets a month, roughly 800 input tokens and 100 output tokens each. That is 80M input and 10M output tokens: $12.00 input plus $5.00 output, so $17 per month. Set reasoning_effort to low and the output side shrinks further.
A coding assistant. 500 sessions a day, roughly 15,000 input tokens (a lot of it repeated file context) and 2,000 output tokens per session. Monthly that is 225M input and 30M output. Without caching: $33.75 plus $15.00, so $48.75. With 70% of input hitting cache: input drops to roughly $14.85, bringing it to about $30.
A document pipeline with images. 10,000 documents a month, roughly 40,000 tokens each including image content, and 1,500 output tokens. That is 400M input and 15M output: $60.00 plus $7.50, or $67.50 per month for work that requires vision.
None of these are large numbers. That is the point of this model.
Cached input is the biggest lever
At $0.03 per million against $0.15 for fresh input, cached tokens cost one fifth of uncached ones. Any workload with a stable prefix, a system prompt, a document being asked about repeatedly, a codebase in context, should be structured so that prefix stays identical across calls.
The mistake that quietly destroys cache hit rates is putting anything variable near the front of the prompt. A timestamp, a request id, or a user name in the system prompt invalidates everything after it. Put the stable content first and the variable content last.
Z.ai has also listed cached-input storage as free for a limited period. Treat that as promotional and verify current terms rather than building an architecture that depends on it.
The second lever is reasoning effort
reasoning_effort defaults to max on this model. That is the most expensive setting, and it is what you get if you never touch the parameter.
The three modes are low, high, and max. Reasoning tokens are billed as output, so a task that does not need deliberation is paying output rates for thinking nobody reads. For classification, extraction, routing, and formatting, low can cut the output side substantially.
The setup is covered in our API guide. It is a one-line change and it is the first thing to try if your bill looks higher than the math above suggests.
How it compares
Against its own sibling, at list price:
| Model | Blended per 1M |
|---|---|
| GLM-5.3-Flash | $0.10 |
| GLM-5.3 | $0.90 |
Roughly a nine-fold gap for a three-point difference on the Artificial Analysis Intelligence Index, 57 against 60. Our full comparison covers when that trade is wrong, and the short answer is: when you stream long responses to a waiting human, because GLM-5.3 generates almost twice as fast.
Against GLM-5.2, Z.ai’s claim is roughly one tenth the price, alongside a cost-per-task figure of $0.045. Our GLM-5.2 pricing breakdown has the older rate card if you are budgeting a migration.
Against the cheap multimodal tier from other vendors, GLM-5.3-Flash is competitive on price and unusual in being MIT-licensed, which means the self-hosting option stays open. Our Gemini 3.7 Flash pricing post covers the closest comparison on the Google side.
Reseller prices differ, sometimes a lot
The model is available through Z.ai directly plus OpenRouter, Cloudflare Workers AI, Vercel AI Gateway, DeepInfra, Novita, GMICloud, Baseten, and io.net.
These do not all charge the same. AIHubMix, for example, has listed rates around $0.113 input and $0.394 output, which sits between Z.ai’s promotional and list pricing. Some providers pass through the launch discount and some do not.
If you are routing through an aggregator, check its current rate rather than assuming it mirrors Z.ai. The convenience of a unified gateway sometimes carries a margin, and at these absolute price levels a percentage markup is easy to overlook and easy to tolerate.
When self-hosting starts to win
The weights are MIT-licensed, so running it yourself is a real option, and the break-even math got harder for self-hosting when the API price dropped this low.
A rough frame: full-precision serving wants an 8x H200-class node, which runs somewhere in the region of $24 to $48 per day on rented cloud capacity. Call it $1,000 per month as a fixed cost that accrues whether the node is busy or idle.
At list API pricing, $1,000 buys roughly 6.7 billion input tokens. You need to be running enormous, sustained volume before the fixed cost of a node beats that. For most teams the honest answer is that self-hosting GLM-5.3-Flash is about control, data residency, or licensing rather than cost.
The exception is the quantized tier. GGUF quantizations exist, and running a quantized model on hardware you already own changes the math completely because the marginal cost is electricity. Our local-run guide covers what each hardware tier can actually do.
The coding plan alternative
If your usage is a developer working in Claude Code or Cline rather than an application making API calls, per-token pricing may be the wrong model entirely. The GLM Coding Plan starts around $18 per month, includes GLM-5.3-Flash, and reportedly grants three times the usable quota of GLM-5.3. Z.ai’s documentation also mentions off-peak calls consuming half the standard points.
For a single developer coding daily, a flat plan is usually cheaper and always more predictable than metered API access. Our harness setup guide covers wiring it up, and our comparison of coding plans puts the GLM plan against the alternatives.
Watch the output side
Input pricing gets the attention because the numbers are bigger in volume. Output is where budgets actually break, for two reasons.
Output costs more than three times what input costs per token, at $0.50 against $0.15. And reasoning tokens are billed as output. A model set to max reasoning effort can generate several times more tokens than appear in the final answer, all of it charged at the output rate.
That combination means an application with a modest prompt and a chatty model can end up output-dominated even though it looks input-heavy on paper. The blended $0.10 per million figure assumes a 7:2:1 ratio that many real workloads do not match.
The fix is measurement rather than estimation. Every completion response carries a usage object with the token counts for that call. Log it, aggregate it, and compare the real ratio against the assumption you budgeted on. If output is running above about 15% of your total tokens, reasoning_effort is the first place to look and prompt length is the second.
What to do before September 9
Three things worth doing while the discount is live:
- Measure your actual token mix. The blended rate assumes 7:2:1. If yours is output-heavy, your effective cost is closer to the $0.50 output rate than to the $0.10 blended figure.
- Check your cache hit rate. At a 5x price difference between cached and fresh input, this is where the money is.
- Re-run your cost model at list price. Everything doubles on September 10. If the workload only pencils out at the promotional rate, you have a problem to solve now rather than in two weeks.
Tracking real token usage means capturing real responses, since the usage object on each completion carries the input, output, and cached-token counts you need. Running your representative calls as a saved collection in Apidog, with the model id as an environment variable, gives you those numbers per request and lets you re-run the same suite against GLM-5.3 to compare the bills rather than the marketing.
FAQ
Is GLM-5.3-Flash free? No. It was free for about a week while it ran anonymously as “ox-alpha” before launch, but that ended on announcement. The current 50% discount is not the same as free.
When exactly does the discount end? September 9, 2026. List pricing applies from then.
Why do different sites quote different prices? Some quote the promotional rate, some quote list, and resellers set their own margins. Always check the provider you actually call.
Does image input cost extra? Images consume context tokens and are billed as input. There is no separate image surcharge, but a high-resolution image consumes a meaningful number of tokens.
Is it cheaper to self-host? Only at very large sustained volume, or on hardware you already own using a quantized build. At $0.15 per million input tokens, renting an 8x H200 node is hard to justify on cost alone.
