Billing & Credits
How Critique allocates credits across plans, what consumes them, and how to avoid surprise spend.
Critique uses credits as a single currency for AI workloads: automated PR reviews, Remedy, chat, repository indexing, and related features. Your dashboard Usage page is the best place to see remaining balance and history.
Plans and model access
Plans (Standard, Pro, Ultra) control two things:
- Monthly credit pool — How much compute your organization can use in a period.
- Model catalog access — Higher plans unlock more capable (and typically higher–credit-floor) models in review and Remedy policy pickers.
See Pricing for list prices and plan positioning. The Models page shows which models exist and how they are tiered.
Billing period
Allowances reset on a calendar-month cycle (UTC) unless your contract specifies otherwise. Bonus credits from referrals or promotions are added on top of the monthly pool; check Usage to see how much bonus balance you still have.
What consumes credits
| Surface | Typical cost profile |
|---|---|
| PR review | Highest — multi-stage pipeline (evidence gathering, parallel specialists, synthesis) |
| Remedy | High — sandbox execution with validation loops |
| Chat | Variable — light Q&A is modest; tool-heavy threads cost more |
| Builder jobs | High — cloud agent runs |
| Inference API | Variable — per-token metering on hosted models (/api/v1/chat/completions); see Inference API |
| Indexing / embeddings | Background — refreshes repository memory; can spike after large installs |
Remedy and full reviews cost more than a single @critique question on a PR, which uses one focused model call.
Credit floor (runaway protection)
Each model in the catalog has a credit floor: the minimum balance Critique expects before starting a run with that model. If your remaining credits are below a model’s floor, Critique will not start an expensive run with that model — it may fall back to a cheaper allowed model (per policy) or pause until the next period.
This protects teams from one enormous diff draining the entire month on a frontier model.
PR review vs chat
Credits are one pool, but Usage breaks down spend by surface so you can see whether reviews or chat drove a spike. Automated reviews do not “borrow” from a separate chat wallet — planning should assume both draw from the same allowance.
Inference API
v5.2 adds OpenAI-compatible inference billed from the same credit pool. Token usage on deepseek/deepseek-v4-flash, tencent/hy3, and nvidia/nemotron-3-ultra-550b-a55b converts to credits at published per-million-token rates on /inference-api. Hy3 review runs bill at 1 credit per run on the PR review catalog.
Optional per-user caps (daily/monthly credits, request limits, credit reserve for review) live under Settings → Connections. Signed-in usage charts: /inference-dashboard.
DeepSeek V4 Flash offers a training opt-in tier at 75% off token rates when you explicitly allow prompt logging — account toggle or X-Critique-DeepSeek-Training-Opt-In: true. See Inference API.
Bring your own key (BYOK)
Teams can add an LLM Gateway, OpenRouter, and/or CrofAI (nahcrof) API key in Settings and use the $8/month BYOK harness ($8 USD · €8/mo EUR) so model tokens bill the provider directly while Critique runs orchestration. Usage shows externally billed model calls separately from bundled credits. When multiple BYOK keys are saved, LLM Gateway takes precedence, then Crof, then OpenRouter.
See Bring your own key (BYOK) for setup, what surfaces use your key, and how review gates change. Execution agents (Cursor, Codex, etc.) are covered under BYOA.
When reviews pause
If credits are exhausted before the period resets:
- New automated PR reviews may be queued or skipped until balance returns (or BYOK applies).
- Remedy and heavy builder jobs generally require a positive balance.
- Chat may still allow limited interaction depending on plan and configuration — check Usage for the current gate.
There is no published per-token formula on this site; consumption is derived from actual model usage and catalog floors.
Tips for teams
- Start with default policy and Standard; upgrade plan or models only when findings quality plateaus.
- Use Checkpoint to avoid reviewing obvious low-trust PRs.
- Use
@critiquefor targeted questions instead of/reviewwhen you do not need a full pipeline rerun. - Monitor Usage weekly after onboarding a large monorepo.