Billing
Each organisation has prepaid credit, shared by its projects. New organisations start with $5 of trial credit. Your balance and recent charges are under Settings → Usage.
What costs what
| What | Price |
|---|---|
| On-demand runs (each in a fresh sandbox) | $0.05 per sandbox hour, by the second |
| Instances and fleet members | By uptime, from $0.05 per hour for lgr.small to $0.40 for lgr.xlarge, charged every minute and when they stop |
| Model tokens on platform models | List price for the model; see below |
| Model tokens on your own key | Nothing on LagrangeCloud; your provider bills you |
| Improvement replays | Nothing; LagrangeCloud pays for them |
Runs on an instance aren't charged sandbox time on top of its uptime. Calls through an agent's endpoint, scheduled runs and webhook runs are runs like any other. Callers never pay you or LagrangeCloud directly: everything is charged to the organisation that owns the agent.
Platform models and your own key
Runs call models through the platform unless the project has its own key for
the harness's provider: OPENAI_API_KEY for Codex, ANTHROPIC_API_KEY for
Claude Code and Lagrange, set under Settings → Secrets. With your own key,
only sandbox time is charged.
Platform tokens are charged per million tokens:
| Model | Input | Cached input | Output |
|---|---|---|---|
tensorblock/gpt-6.1-sol | $2.00 | $0.10 | $10.00 |
tensorblock/gpt-5 | $1.25 | $0.125 | $10.00 |
tensorblock/claude-opus-5-5 | $4.00 | $0.20 | $20.00 |
tensorblock/claude-sonnet-5-5 | $2.00 | $0.20 | $10.00 |
tensorblock/claude-fable-5-1 | $10.00 | $0.25 | $50.00 |
The model picker in the agent's Settings shows what this deployment offers.
Running out
When the credit runs out, new runs, messages, sessions and instances are
refused with HTTP 402 insufficient_credit, and running instances are
stopped once they are idle. Their disks are kept, so they can be started
again later. Drafting an agent and Adjust also need some credit left.
Credit can't be bought in the console yet.
Limits per run
Set a time or cost limit per run in the agent's Settings, or limits
in lgr.yaml:
limits:
max_minutes: 30
max_cost_usd: 2A run that reaches its time limit stops and is marked failed. The cost limit is enforced by Claude Code and Lagrange, which also cap it at the credit left; Codex runs have no hard cost cap yet, so give them a time limit.