Pricing
A tier is a monthly allotment of weighted tokens and a published concurrency limit. Nothing is metered that is not on this page, and every figure here is read from the same table the gateway bills against.
Start with 100K weighted tokens, free.
Roughly 2M input tokens on GLM-5.3 Flash. A one-time credit on every new account, before any tier and without a card. Pick a tier when the credit runs out, not before.
200K if you arrive through someone else's invite link.
Held back on GitHub accounts younger than 14 days, which is an anti-abuse measure.
Three tiers, billed monthly
Allotments are sized so that a tier burned to the last token still costs us less than it earns. That constraint is why the allotment is a real number rather than the word unlimited.
- Monthly allotment
- 5.3M WT
- Concurrent requests
- 4
- Effective rate
- $2.83/MWT
- Monthly allotment
- 13.8M WT
- Concurrent requests
- 6
- Effective rate
- $2.83/MWT
- Monthly allotment
- 35M WT
- Concurrent requests
- 8
- Effective rate
- $2.83/MWT
These are introductory rates. They are the rates early subscribers pay, and if they later rise, the terms you subscribed on are the terms you keep: the price table is versioned, and a change opens a new row rather than rewriting the one you agreed to.
What the concurrency number means
It is the number of requests your key can have in flight at the same time. A request beyond it waits its turn; it is not dropped, and it does not cost extra.
This is a fairness limit between customers, not a throttle that tightens when we are busy. It does not change with load, it is not lowered at peak, and it is the same number at three in the morning as at any other hour. An agent that runs a handful of tool calls in parallel wants more of it; a single chat session never notices it.
Credit packs, bought once
Credits are spent only after the monthly allotment is gone, and they never expire. They exist so that an unusual month does not have to become a bigger subscription.
| Pack | Credits | Price | Rate |
|---|---|---|---|
| Starter | 2M WT | $9 | $4.50/MWT |
| Standard | 8M WT | $32 | $4.00/MWT |
| Bulk | 25M WT | $89 | $3.56/MWT |
A pack costs more per token than any subscription. That ordering is deliberate and it is stated here rather than discovered later: if you are buying packs every month, a larger tier is cheaper, and we would rather you knew that than not.
What a weighted token is
One weighted token is one input token on the reference model. Everything else (output, cached input, other models) converts into that unit at a published rate, so one allotment covers the whole catalog without a per-model price list.
WT = (in × W_IN + cached × W_CACHED + out × W_OUT) × M_model
W_IN = 1.0 # definitional anchor: one input token on glm-5.3
W_CACHED = 0.1857 # what a cache hit costs against a miss
W_OUT = 3.1429 # the provider's output ÷ input price ratio
M_model = 1.0 # glm-5.3 is the reference; others convert to itEvery response carries the arithmetic behind its own charge: the token counts, the weights applied, and the multiplier, so the formula above is checkable per request rather than taken on trust. The per-model weights and a worked receipt are on /models.
What an allotment buys, per model
A weighted token is not a token, and the difference is large enough to matter: the same balance buys an order of magnitude more of the cheapest model than the dearest. Below is Tier 1’s 5.3M WT, converted.
| Model | Multiplier | 5.3M WT buys, in input tokens |
|---|---|---|
| glm-5.3-flash | ×0.05 | 98.9M |
| deepseek-v4-flash | ×0.31 | 16.9M |
| deepseek-v4-pro | ×0.94 | 5.6M |
| glm-5.3 · reference | ×1.00 | 5.3M |
| kimi-k3 | ×2.14 | 2.5M |
These are input tokens, on each model separately. Output tokens weigh more than input (between two and five times, depending on the model), so a real workload lands below these figures by an amount that depends on how much it generates. There is no single tokens number that is true across the catalog, which is why this table has a row per model rather than a headline.
A multiplier is not a markup. It is what the model costs to serve relative to the reference, so a cheaper model genuinely makes an allotment go further rather than earning us more.
What happens when the allotment runs out
The allotment goes first, then any credits you hold. At zero, requests stop with a 402 and a message naming your reset date. Nothing is served that has not been paid for, and nothing is charged for a request that was not served.
A hard stop is the honest behaviour but not the comfortable one: it lands mid-run, at whatever hour your agent happens to be working. Buying a credit pack resumes service immediately. We would rather stop you than quietly run up a bill you did not agree to.