Next batch: Reservations open, closes Nov 20

Private GLM-5.3 capacity for your agents.

Reserve guaranteed slots on GPUs we control, allocated to a limited number of teams per server and never resold by the token. Each slot is one agent working at full speed, for a flat monthly price instead of a per-token bill. Your prompts, code and model outputs are processed only in the EU, and never stored.

From $900 per month. Nothing is charged until your batch confirms.

Individual developer? Use the pay-as-you-go API, with free credit to start.

What you get

Capacity that is yours, not a shared queue.

Guaranteed concurrency

A slot is one agent request running at full speed, any time during your hours. Not a rate limit you share with strangers, and not a queue behind someone else's batch job.

A flat monthly price

No per-token bill that grows with every long session. Your agents can read the whole repo, retry, and think for as long as they need. The price does not move.

EU only, never stored

Your prompts, code and model outputs are processed only in the EU, and never stored. No Chinese endpoint in the path, and a person who answers when you write.

The maths

A busy agent is expensive per token.

Coding agents re-read large contexts on every turn. At trusted per-token prices that adds up fast. Move the sliders to your own hours.

Per-token, trusted providers (est.)
$2,864 /mo
One 90K business-hours slot
$900 /mo
Difference
$1,964 /mo

Estimate for one coding agent working nonstop during those hours: about 44K tokens of context per turn, mostly cached, roughly 1K tokens out, at GLM-5.3 prices from trusted US providers (about $1.40 in and $4.40 out per million tokens, cached input discounted). Agents that idle part of the time cost less per token. If yours mostly idle, per-token may suit you better, and we will say so.

Measured

Fast with every slot busy.

These are measured on the hardware your slot runs on, with the box full, not a single request on an idle machine.

1.0 s
median time to first token
p95 1.71 s
50 tok/s
per agent, 90K slots
24 tok/s with every 45K slot busy
96%
of each prompt reused, not reprocessed
so long agent sessions stay fast
100%
tool calls parsed
zero errors at every load level

Measured 2026-10-06 on the hardware your slot runs on, with a real coding-agent loop over a real codebase: tool calls every turn, sessions compacting as they grow. Speeds are per agent with the box at its full slot count, not a single request on an empty machine.

How it works

Reserve now, pay when your batch confirms.

Slots are allocated in batches, so every server runs fully allocated. That is what keeps the price low.

1

Reserve

Pick a context size, 24/7 or business hours, and how many slots. Nothing is charged, and you can cancel for free until the batch confirms.

2

The batch confirms

The next batch starts on Tuesday, December 1, 2026, or about 7 days after it fills if that comes first. We email you before anything is invoiced. The invoice follows, due in 7 days.

3

You're live

You get an API key for your private GLM-5.3 endpoint. It works with any OpenAI-compatible tool. Invoices come on the 1st of each month after that.

Pricing

One price per slot, per month.

90K-context slots from $900 a month for business hours, or $2,400 for 24/7. 45K and 128K sizes too, and up to 15% off for paying ahead.

Just want per-token API access?

For individual developers and smaller workloads, our pay-as-you-go API serves GLM-5.3 and other open models through one OpenAI-compatible endpoint, with free credit to start.