Search across all documentation pages
OpenRouter bills most gateway usage through USD-denominated credits: you deposit funds, inference burns credits at pass-through provider rates, and a fee applies when you buy credits (not as a hidden markup on each token).
BYOK is a parallel path: your provider is charged for tokens, while OpenRouter may take a percentage fee after a free monthly request allowance.
All dollar thresholds, percentages, and free-tier numbers below are verify at build against current OpenRouter FAQ and pricing pages.
Credits are deposits in USD terms on your OpenRouter account.
When a paid request completes, OpenRouter calculates cost from the provider's token (or request/image/reasoning) pricing and deducts that amount from your balance.
Docs state no markup on underlying provider inference prices; the business model centers on fees when purchasing credits (and BYOK fees after free requests).
Typical flow:
Refunds, when available, are time-bounded (FAQ has described unused-credit refund windows measured in hours, with fees non-refundable - verify at build). Crypto purchases may be non-refundable.
Terms may allow expiry of unused credits after a long idle period (e.g. on the order of one year - verify at build).
| Cost component | Paid via credits | Notes |
|---|---|---|
| Prompt tokens | Yes | Per model list price |
| Completion tokens | Yes | Often different $/M from prompts |
| Reasoning / internal tokens | Sometimes | Model-specific |
| Image / request fees | Sometimes | Shown on model pages |
| Credit purchase fee | At top-up | Not per token |
| BYOK % fee | From credits after free quota | Provider still bills you |
Always read the model card: agent workloads with large tool observations spike prompt tokens fastest.
Two layers can stop spend:
limit, limit_remaining, reset policy via key APIs).Agents in staging should use keys with low ceilings so a runaway loop cannot drain the org balance.
import os
import httpx
r = httpx.get(
"https://openrouter.ai/api/v1/key",
headers={"Authorization": f"Bearer {os.environ['OPENROUTER_API_KEY']}"},
timeout=30.0,
)
r.raise_for_status()
print(r.json()) # usage, limit_remaining, is_free_tier, etc.Use this in boot checks or cron, not on every token, unless you accept the extra call volume.
| Code | Meaning | Operator action |
|---|---|---|
| 402 | Payment required - insufficient credits or key cap | Add credits / raise key limit |
| 429 | Rate limited (platform free caps or upstream) | Backoff; change model; buy credit threshold for free tier |
| 401 | Bad or disabled key | Rotate secrets |
Do not treat 402 as "retry forever." Surface a budget stop condition in the agent.
With Bring Your Own Key:
So BYOK still needs a small credit balance once you exceed the free request threshold, even if providers are paid elsewhere.
Free :free variants do not burn inference credits the same way paid models do, but:
The Activity view (and exports/APIs) break down usage by model, provider, and API key.
Finance teams should reconcile:
Estimate cost per successful task, not cost per call.
A single support agent run might be 8 model calls with growing context. Instrument:
| Pattern | Billing setup | Risk control |
|---|---|---|
| Startup single account | Shared credits, env-specific keys | Tight staging key limits |
| Agency multi-client | Separate keys or workspaces per client | Per-key caps = client budgets |
| Enterprise | Credits + BYOK for strategic providers | Privacy filters + procurement review |
| Eval lab | Dedicated key, auto top-up off | Hard monthly cap |
| Approach | Strength | Weakness | Best fit |
|---|---|---|---|
| Credits only | Simple, full catalog | Purchase fee; balance ops | Most agent teams |
| BYOK heavy | Uses cloud commits; high control | Dual bills; fee after free requests | High volume on few providers |
| Direct only | Possible enterprise discounts | No unified long-tail models | Single-vendor products |
| Hybrid | Flexibility | Accounting complexity | Mature multi-model platforms |
Primarily via fees when you buy credits and via BYOK fees after free request allowances. Confirm current policy at build.
Usually no. Model pages list separate rates. Agent traces should store both.
USD-denominated pricing on the site and API.
Policies are time-limited and method-dependent. Check the live FAQ and Credits page; do not assume open-ended refunds.
Terms have reserved rights to expire unused credits after a period such as one year - verify current terms.
Often negative balance, key limits, or account state issues. Add credits or inspect key limits via GET /api/v1/key.
Issue a dedicated API key with a credit limit and optional reset interval; enforce max turns in the agent loop as a second brake.
Payment processors may add taxes depending on location. Check checkout and invoices.
Billing is usage-based on tokens (and other metered units), not on streaming itself. Failed mid-stream calls may still consume partial usage depending on provider accounting.
Docs describe it as a percentage of what the same model/provider would cost on OpenRouter, deducted from OpenRouter credits after the free monthly request count.
Only while BYOK stays within free request thresholds and you do not need credit-funded models. Beyond that, keep credits available for fees and non-BYOK routes.
Activity dashboard filtered by key/model/provider; export as needed for finance.
No. Sample periodically or on 402. Cache remaining limit in your control plane.
If you buy credits in large chunks, the fee amortizes across many tokens. Tiny top-ups maximize fee drag - batch purchases when cash flow allows.
Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana
@solana/web3.js+@solana/spl-token.
Reviewed by Chris St. John·Last updated Jul 16, 2026