Budgets and tiers
The credit model, per-task budgets, and plan limits
Budgets and tiers
Foxl Code meters everything an agent spends - both LLM tokens and the cloud compute its VM uses - against a single credit balance, with a hard ceiling per task so a run can never overshoot.
Credits
Both LLM usage and AgentCore compute draw from one pool of credits.
- 1 credit = $0.04 USD.
- A typical 30-minute Sonnet task lands around 7-8 credits.
Per-task budgets
Every task carries a spending limit and the relay tracks cost in real time:
- Auto-estimate - computed up front from the model, expected duration, and the VM's vCPU/memory profile.
- Custom budget - set a ceiling before launch in the composer's "Budget (USD)" field. It must be at least the estimate.
- Multiplier - a per-user scaling factor in Settings (clamped between 0.5x and 2x) for repos that routinely need more headroom.
- Default per-task budget - an account-level floor in Settings (default $500) applied as the minimum on the auto budget for every task; the per-task "Budget (USD)" override on the new-task dialog still wins.
Budgets are a hard cap, enforced live. The moment a task's cumulative spend reaches its ceiling, the relay closes the upstream connection - the agent commits what it has and stops. There are no overage charges.
Tiers
| Tier | Price | Credits / mo | Concurrent agents | Models |
|---|---|---|---|---|
| Free | - | 10 | 5 | Codex (GPT) + Cursor (BYOK) only; Claude models require Pro |
| Pro | $20 / mo | 500 | Unlimited | All, incl. Opus + Codex |
| Ultra | $200 / mo | 10,000 | Unlimited | All + priority |
| Enterprise | Usage-based | Unlimited | Unlimited | All + dedicated infra |
LLM rates
Token prices (USD per million tokens) used by the estimator and meter:
| Model | Input | Output | Cache read |
|---|---|---|---|
| Claude Fable 5 | $10 | $50 | $1.00 |
| Claude Opus | $5 | $25 | $0.50 |
| Claude Sonnet | $3 | $15 | $0.30 |
| Claude Haiku | $1 | $5 | $0.10 |
| GPT-5.5 | $5 | $30 | $0.50 |
| GPT-5.4 | $2.50 | $15 | $0.25 |
Compute is billed separately from tokens but out of the same credit pool: AgentCore charges for active vCPU time plus memory for the life of the VM. Because coding agents spend most of their wall-clock waiting on I/O, the active-vCPU portion is usually a small fraction of the total.