Quick answer: GLM 5.3 Flash lists at $0.15 per 1M input tokens, $0.03 per 1M cached input, and $0.50 per 1M output tokens. The $0.075 / $0.25 figure quoted almost everywhere is that list price with a temporary 50% launch discount applied, currently honoured by Z.AI's own endpoint, Novita, and GMICloud. Six other providers already charge full list. Budget at $0.15 / $0.50 and treat the discount as upside, not baseline.
Here is the trap this article exists to defuse. You search "GLM 5.3 Flash cost," you find $0.075, you build a forecast on it, and three months later the promotional window closes and your inference bill doubles with no code change and no warning. It is not a hypothetical — it is what a launch discount does by design.
The numbers below come from Z.ai's published API rates and from OpenRouter's public endpoints API, which exposes every provider's live per-token price for a given model. We queried that API on August 27, 2026 and reproduce the full ten-provider table rather than a single headline rate, because "the price of GLM 5.3 Flash" is not one number — it is a spread. We run cost-sensitive inference daily, so the worked examples below are shaped by how token bills actually accumulate, not by spec-sheet arithmetic.
What This Article Solves
The pain point: you cannot forecast spend from a headline rate. Three things break the naive calculation for GLM 5.3 Flash specifically — a promotional discount that is not labelled as promotional in most coverage, a 2x price spread across the providers serving the same weights, and a cached-input rate that changes the economics of long-context work by an order of magnitude if you use it and not at all if you do not.
You will leave with the list price, the discount price, the provider table, three worked cost scenarios, and a cache-hit calculation that most pricing pages omit entirely.
The List Price (Use This One)
Z.ai's published API rates for glm-5.3-flash:
| Direction | Price per 1M tokens |
|---|---|
| Input | $0.15 |
| Cached input | $0.03 |
| Output | $0.50 |
Two things stand out against the class. First, the input:output ratio is 1:3.3, steeper than the 1:2 that cheap models usually run. Output volume is the lever that moves your bill here — capping max_tokens and discouraging preamble buys you more than shortening prompts does.
Second, the cached input rate is 5x cheaper than fresh input. On a model with a million-token context window that is not a rounding detail; it is the whole ballgame for repository-scale and document-scale work. More on that below.
The Discount — and Exactly Who Is Honouring It
OpenRouter's model page shows GLM 5.3 Flash with a 50% discount applied, which produces the $0.075 / $0.25 / $0.015 numbers in wide circulation. That discount is not universal. Here is every provider serving the model, straight from OpenRouter's endpoints API on August 27, 2026:
| Provider | Input / 1M | Output / 1M | Cache read / 1M | Context | Max output |
|---|---|---|---|---|---|
| Z.AI | $0.075 | $0.25 | $0.015 | 1,048,576 | 131,072 |
| Novita | $0.075 | $0.25 | $0.015 | 1,048,576 | 131,072 |
| GMICloud | $0.075 | $0.25 | $0.015 | 1,048,576 | 943,718 |
| Venice | $0.09375 | $0.3125 | $0.01875 | 1,048,576 | 131,072 |
| Modal | $0.14999 | $0.49995 | $0.03 | 1,048,576 | 943,718 |
| Parasail | $0.15 | $0.50 | — | 1,048,576 | 943,718 |
| DeepInfra | $0.15 | $0.50 | $0.03 | 1,048,576 | 943,718 |
| Baseten | $0.15 | $0.50 | $0.03 | 1,048,576 | 131,072 |
| Cloudflare | $0.15 | $0.50 | $0.03 | 1,310,720 | 1,179,648 |
| Io Net | $0.15 | $0.50 | $0.03 | 262,144 | 131,072 |
Read that table twice, because it contains three separate decisions.
The price spread is 2x for identical weights. Every fp8 endpoint here is serving the same MIT-licensed model. If your router picks by availability rather than price, you are paying double at random.
The context ceiling is not uniform. Io Net serves 262,144 tokens — a quarter of the advertised window. If you route long-context jobs by cheapest-available and land on Io Net, you will get truncation errors, not a bill surprise. Cloudflare, conversely, lists 1,310,720.
Max output varies by 9x. 131,072 on Z.AI and Baseten; 943,718 on DeepInfra, Parasail, GMICloud and Modal; 1,179,648 on Cloudflare. For long-generation jobs this determines whether a request completes in one call or needs continuation logic.
Independent tracker Artificial Analysis publishes a blended price of $0.10 per 1M tokens for GLM 5.3 Flash, using a 7:2:1 cache-hit / input / output mix. That is a good cross-model comparison number and a bad budgeting number — unless your traffic genuinely hits cache 70% of the time.
Worked Cost Examples
All figures below use the list rate ($0.15 in / $0.50 out), because that is what you should be forecasting on. Halve them if the discount is live on your provider and you are comfortable with the risk.
Example 1 — A coding agent session. One serious agentic task: ~120K input tokens (repo context, tool results, iterations) and ~15K output tokens. Input: 0.12M x $0.15 = $0.018. Output: 0.015M x $0.50 = $0.0075. Total ≈ $0.026 per task. Roughly 38 substantial agent runs per dollar.
Example 2 — High-volume chat. 100,000 requests at ~2K input and ~500 output each. Input: 200M x $0.15 = $30.00. Output: 50M x $0.50 = $25.00. Total = $55.00 per 100K requests. Note that output is 45% of the bill on only 20% of the tokens — that 1:3.3 ratio compounding.
Example 3 — Long-context document processing. 500 documents, ~400K input tokens each, ~4K output each. Input: 200M x $0.15 = $30.00. Output: 2M x $0.50 = $1.00. Total = $31.00. Cheap — until you notice you are paying full input rate on a shared system prompt 500 times over.
The Cache Rate Is Where the Real Money Is
Example 3 is the one worth re-running with caching. Suppose 60% of each request's input is a stable prefix — a system prompt, a schema, a code style guide, a fixed corpus of reference documents:
- Cached portion: 120M tokens x $0.03 = $3.60
- Fresh portion: 80M tokens x $0.15 = $12.00
- Output unchanged: $1.00
- New total: $16.60 — a 46% reduction with zero change to model or quality.
This is the highest-leverage optimisation available on GLM 5.3 Flash, and it is structural rather than clever: put your stable content first and keep it byte-identical between requests. Note that Parasail does not publish a cache-read rate at all, so cache-heavy workloads should not route there. If you want to see what your own prompt shape costs before wiring anything up, run a GLM 5.3 Flash chat on glm5.app and watch the token counts on a realistic prompt.
How the Price Compares
The comparison that Z.ai itself leads with is against its own previous generation: GLM 5.3 Flash is reported to beat GLM-5.2 across benchmarks and real workloads at roughly one-tenth the price. That is the strongest claim in the launch and it is a within-family, same-methodology one, which makes it more trustworthy than cross-vendor charts.
Against the obvious open-weight rival, DeepSeek V4 Flash lists at $0.14 input / $0.28 output — cheaper on output by a wide margin, more expensive on Artificial Analysis's blended metric ($0.23 vs $0.10 per 1M) because DeepSeek's cache discount is less aggressive. We compare the two in full in GLM 5.3 Flash vs DeepSeek V4 Flash. Which one is cheaper for you depends entirely on your cache hit rate and your input:output ratio. There is no universal answer, and any article that gives you one is guessing.
The differentiator to hold onto: most GLM 5.3 Flash pricing coverage quotes one discounted number and stops. The three facts that actually determine your bill are (1) the discount is temporary, (2) six of ten providers already ignore it, and (3) the 5x cache differential outweighs provider shopping for any workload with a stable prefix. Optimise in that order.
Coding Plan vs. Pay-Per-Token
If your usage is a developer with an IDE rather than a service with traffic, the GLM Coding Plan changes the arithmetic. GLM 5.3 Flash is available across the Lite, Pro, and Max tiers, and carries roughly 3x the usable quota of GLM-5.3 on the same plan. For steady daily coding use, a subscription tier is generally cheaper than metered tokens; for bursty or programmatic traffic, pay-per-token wins because you are not buying idle capacity.
The decision rule is simple: if you can predict your monthly token volume within about 2x, price both and pick. If you cannot, start metered — you can always move to a plan once you have a month of real numbers.
Either way, the cheapest sanity check is free: run a GLM 5.3 Flash session on glm5.app on a task that looks like your real traffic, count the tokens it actually burns, and put that number into the arithmetic above before you commit to a plan or a forecast.
Frequently Asked Questions
Is GLM 5.3 Flash free? Not any more. It was free during its preview week on OpenRouter under the codename Ox Alpha. Today it is paid on every hosted endpoint. The weights are free in the open-source sense — MIT-licensed on Hugging Face — but self-hosting means supplying roughly 328 GB of GPU memory, which is not free in any practical sense.
Why do I see $0.075 and $0.15 for the same model? $0.15 is the list price; $0.075 is a temporary 50% launch discount currently honoured by Z.AI, Novita, and GMICloud. Both are real, and one of them is going away.
How do I pin a provider so I actually get that price?
Through OpenRouter's provider.order block — the syntax is in our GLM 5.3 Flash on OpenRouter guide.
Which provider is cheapest? As of August 27, 2026: Z.AI, Novita, and GMICloud tie at $0.075 / $0.25. But check the context ceiling before routing — GMICloud's max output is 943,718 while Z.AI's is 131,072, and Io Net serves only a 262K context.
How does the cached input rate work? Repeated prefix content bills at $0.03 per 1M instead of $0.15 — 5x cheaper. It applies to the stable, byte-identical head of your prompt. Structure prompts with fixed content first and variable content last to maximise the hit rate.
Is it cheaper than DeepSeek V4 Flash? On list output tokens, no ($0.50 vs $0.28). On Artificial Analysis's blended 7:2:1 metric, yes ($0.10 vs $0.23). The answer flips based on your cache hit rate, so run your own numbers rather than trusting either headline.
Sources
- GLM 5.3 Flash on OpenRouter — model page showing the 50% launch discount and provider list.
- OpenRouter endpoints API for z-ai/glm-5.3-flash — the ten-provider price, context, and max-output table, queried August 27, 2026.
- zai-org/GLM-5.3-Flash — Hugging Face model card — MIT license, model size, and the one-tenth-the-price-of-GLM-5.2 claim.
- Artificial Analysis — GLM-5.3-Flash — independent blended price ($0.10 / 1M) and performance measurements.
- DeepSeek API pricing — first-party rates for the DeepSeek V4 Flash comparison.
Prices verified August 27, 2026. Launch discounts expire without notice and provider rates change frequently — confirm on the provider's own page before committing a budget.
