Plans
Counted in requests, not in tokens you cannot see.
Two endpoints, one bill. theta is our own model, built for volume — long agent loops, mechanical edits, tool chatter. glm-5.3 is built for the turns that decide the shape of the work, on a million-token context. Both are counted in requests you can predict, with a live window in your dashboard.
Regional pricing is deliberate: India is billed in rupees through Razorpay, everywhere else in dollars through Stripe.
Plan comparison
| Allowance | starter$10/monthor ₹1,000 in India | proUnlimited theta$25/monthor ₹2,500 in India |
|---|---|---|
| theta requestsrolling 5-hour window | 300 | Unlimited |
| theta concurrencyrequests in flight at once | 5 | 1 |
| Extra theta poolmonthly, usable up to 5 concurrent | — | 5,000 requests |
| glm-5.3 requestsrolling 5-hour window | 20 | 50 |
| glm-5.3 token budgetrolling 5-hour window, across all calls | 2M | 5M |
| Endpointsboth plans get both | theta (ours) · glm-5.3 | theta (ours) · glm-5.3 |
| Context window | 200K theta · 1M glm-5.3 | 200K theta · 1M glm-5.3 |
| API compatibility | OpenAI · Anthropic | OpenAI · Anthropic |
| Quota headers on every response | Included | Included |
| Savings calculator & ledger | Included | Included |
Windows roll continuously rather than resetting on a clock, so a burst never lands on a boundary. Need to try before paying? A one-day trial includes both endpoints at reduced allowance.
01 Savings calculator
public · no sign-in
Hold the workload still and change only the plan.
Your month
- input
- 36.0M tok
- output
- 3.6M tok
- glm / 5h
- 4.2 req
Bought direct, at list
$72.12
600 glm-5.3 requests at list rates, no cache discount applied.
On the plan
$25.00/mo
pro — 50 glm-5.3 requests and 5M tokens per 5-hour window, plus unlimited theta requests.
Difference
65%
$47.12 a month, or $565.47 a year.
Does it fit?
This workload averages 4.2 glm requests per 5-hour window — inside the plan’s 50.
02 How we sustain these prices
Three engineering choices, all disclosed.
Prompt-cache engineering
Stable prefixes so cache reads stay high. This is the single biggest lever on what a request costs us, and caching only works if the prefix is stable — which is also why a session stays on one lane instead of being re-routed every turn.
Context management
Your history stays lean without you editing it. Compression happens after metering, so it never reduces the tokens you are charged for.
Skill routing
Within glm-5.3, a capability-and-cost objective decides whether a turn needs the full model or the flash tier. The full tier is capped at a quarter of your glm calls, so it stays available for the turns that actually need it. You always learn which tier answered.
Training on your traffic
We may use API traffic to train our own models. One toggle in your dashboard stops it, and the very next request is not stored. The honest limit of that opt-out is stated in the training disclosure.
03 Reference rates
used for valuation only
| Model | Input / M | Cache read / M | Output / M |
|---|---|---|---|
theta | $0.2 | $0.04 | $0.4 |
glm-5.3 | $1.52 | $0.283 | $4.79 |
In the plans, both endpoints are metered per request. These per-token rates are glm-5.3’s list prices, shown so the calculator can compute a like-for-like figure against going direct — and so the dashboard can put a value on what you used.
A glm-5.3 request also counts against a token budget for its window. That is the one place tokens still matter here: a single call can carry a million-token context, so a request count alone would not bound what it costs to serve you.