How Credits Work
Hermes Haven uses a single credit currency called HeavenCredits (HC) to pay for AI inference. You don't need to think about tokens, API calls, or per-request pricing. Your plan includes a monthly allowance, and you can optionally add more.
This page explains what HeavenCredits are, how euros convert to HC, and how credits are consumed when you use AI models.
What are HeavenCredits?
HeavenCredits (HC) are the single unified currency on Hermes Haven. Every plan includes a monthly HC allowance — no need to juggle multiple credit types or worry about priority order. One balance, one currency, one simple model.
Your plan includes a monthly HC allowance. Unused HC does not roll over at the end of the month. If you need more, you can top up your wallet with euros at any time — those credits never expire.
| Plan | HC included / month |
|---|---|
| Trial | 0 HC |
| Starter (9€) | 1,500 HC |
| Plus (19€) | 6,000 HC |
| Pro (39€) | 20,000 HC |
| Enterprise (119€) | 50,000 HC |
The credits can run out. Your agent can't. When your monthly HC are exhausted, the platform enters economy mode: only Level 0 models (our own GPUs) remain available at no cost. Or top up your wallet to keep using premium models.
How euros convert to HeavenCredits
When you purchase HeavenCredits, euros are converted to HC at a flat rate of 1,000 HC per euro. Top-up credits never expire:
| You pay | You receive | Effective rate |
|---|---|---|
| €10 | 10,000 HC | 1,000 HC/€ |
| €50 | 50,000 HC | 1,000 HC/€ |
| €100 | 100,000 HC | 1,000 HC/€ |
Flat rate: 1€ = 1,000 HC. No volume bonuses — every euro buys exactly the same amount.
What happens when you send a request
Every AI request goes through a reserve → settle → release cycle. This ensures you're never overcharged and failed requests don't cost credits:
- Reserve — Before the request is sent to the model, the platform estimates the maximum cost and holds that amount from your balance. The held credits are temporarily removed but not yet charged.
- Settle (success) — If the request succeeds (HTTP 200), the actual cost is calculated from real token usage. The difference between held and actual is returned to your balance. Only the real amount is charged.
- Settle (partial failure) — If the request fails with a client error (4xx, except 402/429), only the input tokens are charged — you don't pay for output that wasn't generated.
- Release (full failure) — If the request fails with a server error (5xx), rate limit (429), or payment error (402), all held credits are returned. You are not charged for failed infrastructure.
Idempotency: If the same request is retried (e.g. network glitch), the system recognises it and doesn't double-charge. Each request has a unique ID that prevents duplicate billing.
Full example: from request to charge
- You send a chat request — Model: GLM-5.2 (Level 3, rate = 5,138 HC/1M). Estimated max: 2,000 HC held from your HC balance.
- Model responds successfully — Actual usage: 100,000 input tokens + 10,000 output tokens. weighted = 100,000 + (10,000 × 5) = 150,000
- Actual cost calculated — hc = round(150,000 × 5,138 ÷ 1,000,000) = 771 HC
- Balance updated — 2,000 HC held → 771 HC charged → 1,229 HC returned to your balance. Net result: HC balance decreased by 771 HC.
What happens when credits run out?
| Situation | What happens |
|---|---|
| HC balance = 0 | Economy mode: Only Level 0 models (our own GPUs) are available. Premium models are blocked until renewal or top-up. |
| Monthly renewal | Your HC allowance is topped up to your plan's amount. Wallet top-up balance is preserved (it doesn't reset). |
Model pricing — how credits are consumed
Each AI model belongs to a cost level. The level determines how many HeavenCredits your requests consume, measured in HC per million weighted tokens. There are six levels (L0–L5):
| Level | HC/M range | Description |
|---|---|---|
| L0 | 0 | Free (own GPU) |
| L1 | 1–500 | Economy |
| L2 | 501–2,000 | Standard |
| L3 | 2,001–8,000 | Premium |
| L4 | 8,001–30,000 | High-end |
| L5 | 30,001+ | Ultra |
Weighted tokens are not raw tokens. The platform weighs different token types to reflect real compute cost:
weighted = uncached_input + (cached_input × 0.1) + (output × 5) + (reasoning × 5)
Output and reasoning tokens cost 5× more than input because they require active generation. Cached input costs 10× less because it's reused from prior context. Level 0 models run on our own GPUs at no cost — they remain available even in economy mode when your credits run out.
For the full catalogue with per-model details, see AI Models & Cost Levels.
Available models
The platform offers 85 models across 6 cost levels. Level 0 (L0) models are free and run on our own GPU infrastructure. Levels 1–5 (L1–L5) are premium models that consume HavenCredits according to their tier:
| Level | HC/M range | Description |
|---|---|---|
| L0 | 0 | Free (own GPU) |
| L1 | 1–500 | Economy |
| L2 | 501–2,000 | Standard |
| L3 | 2,001–8,000 | Premium |
| L4 | 8,001–30,000 | High-end |
| L5 | 30,001+ | Ultra |
For the full catalogue with all models and rates, see AI Models & Cost Levels.
Bring your own keys = zero credits consumed
If you configure your own API keys (any OpenAI-compatible endpoint), requests routed through those keys do not consume any credits from your Hermes Haven balance.
- Heavy users who have negotiated direct pricing with a provider
- Users who need a specific model not in our catalogue
- Teams with existing enterprise agreements with AI providers
See the full model catalogue and cost levels in AI Models & Cost Levels.