How Hyperbolic pricing, GPU billing, and credits work

Prepaid credits, deposits, and account tiers

Hyperbolic doesn't send you a monthly invoice after the fact. Everything except custom enterprise setups runs through compute credits that you purchase ahead of time.

Credits sit in your account as a positive balance and never expire. You can buy credits whenever you want through the dashboard by clicking Deposit in the upper-right corner, typing an amount, and finishing checkout. The minimum purchase is $5. Because credits match the US dollar at a 1:1 ratio, a $20 deposit gives you exactly 20 compute credits.

Your account tier controls what you can launch. Free Tier personal accounts can only run inference, and you have to verify your phone number under Settings > Plan before those API calls will work. Depositing at least $5 upgrades a personal account to the Pro plan.

You don't need a Pro plan if you work inside an Organization. Organization API keys and Playground calls made while switched into an Organization bill the shared balance of that team directly. Personal API keys always pull from your own balance.

To review where your money went, open the Billing page in the dashboard. It lists every deposit, charge, and credit on your account, including pending charges for running instances that haven't settled yet.

If you're weighing Hyperbolic against alternative infrastructure setups, take a look at our breakdown on how Vast.ai billing works or our guide on RunPod pricing for Pods and Serverless.

GPU instances and storage volume billing

Compute charges hit your balance differently depending on whether you run dedicated hardware or use serverless APIs.

On-demand GPU instances bill hourly from your compute credits. There's no minimum charge for on-demand instances, but you can't spin one up on an empty account. At the moment you click create, your balance must be high enough to cover at least one full hour of runtime across all your pending instances, your running instances, and the new instance you want to start.

Reserved GPU instances follow a different rule: you pay for them up-front rather than by the hour. Hyperbolic Private Cloud operates entirely outside this prepaid structure through a custom contract that bills separately instead of deducting from compute credits.

Storage volumes attached to your machines bill hourly per GB of capacity directly from your credit pool. Disk costs keep accruing every hour as long as the volume exists, whether your GPU instance is running code or sitting idle.

Serverless inference pricing on Flux

Flux is Hyperbolic's serverless inference platform built on Forge, which is their unified compute and orchestration layer.

Inference doesn't charge subscriptions or require ongoing minimum spends once you've made the one-time $5 deposit to reach the Pro tier. Each request deducts from your credit balance based on the model's price per million tokens. Hyperbolic doesn't charge for requests that return an error, which saves you from paying for crashed worker loops or malformed inputs.

Requests can be billed across three separate token rates:

• Input tokens: all tokens sent in your prompt, including the system prompt and conversation history.
• Cached input tokens: prompt tokens served directly from the model's cache. If a model supports prompt caching, these tokens are billed at a lower rate than standard inputs.
• Output tokens: every token the model returns, including thinking and reasoning tokens.

Testing models inside the Hyperbolic Model Playground doesn't give you free passes. Playground calls hit the exact same API and deduct credits at standard rates.

The table below shows current per-million token rates for text models available on Flux:

The Flux API is drop-in compatible with the OpenAI SDK. To make requests, point your client to https://api.hyperbolic.ai/inference/v1 and supply an API key that starts with the sk_live_ prefix. If you have an older key from an earlier account setup, it won't work on Flux; you'll have to generate a fresh one in the console.

from openai import OpenAI

client = OpenAI(
    api_key="sk_live_your_api_key_here",
    base_url="https://api.hyperbolic.ai/inference/v1",
)

response = client.chat.completions.create(
    model="deepseek-v4-flash-0731",
    messages=[{"role": "user", "content": "Ping"}],
)

If your stack calls multiple providers alongside Flux, you can see how to centralize your metrics in our guide on monitoring spend across AI APIs in one place.

What happens when your balance hits zero

Prepaid billing protects you from an open-ended credit card bill, but it creates sharp failure modes if your funds run out mid-job.

If your balance drops to zero or goes negative:

1. Inference requests get denied immediately until you add funds.
2. You can't launch any new GPU instances.
3. Running GPU instances may be terminated.
4. Active workloads get interrupted.
5. Storage Volumes may be terminated after their grace period ends.

That third point is where developers get burned. When an instance is terminated because of a zero or negative balance, data stored directly on that terminated machine may be permanently lost. If you haven't saved checkpoints to an external bucket or persistent storage, an unattended zero balance can wipe out hours of training runs.

Auto top-up mechanics and payment retries

To keep instances from dying unexpectedly, you can configure Auto Top-Up in your dashboard. But setting it up isn't enough; the toggle must be both configured and enabled to actually trigger payments.

Auto Top-Up evaluates your balance at most once every 10 minutes. Because of that 10-minute polling gap, a sudden burst of high-throughput GPU jobs or heavy batch inference can burn through remaining funds before the system triggers a reload. That's why Hyperbolic recommends calculating roughly 24 hours of typical usage and setting both your top-up threshold and your reload amount to that 24-hour figure.

Billing attempts charge your primary payment card first. If that charge fails, it falls back to your backup card. When a transaction declines, Auto Top-Up backs off and retries later, and Hyperbolic sends an alert notifying you of the failure.

Need to force a retry immediately after a card issue? You don't have to wait for the backoff timer to expire. Adding a new card or changing which saved card is marked as your primary or backup card triggers an immediate retry.

Monitoring credit drops and revoking keys

Because Hyperbolic doesn't charge postpaid invoices, tracking your burn rate means watching your credit balance drop over time.

Costbrake monitors your spending across AI, cloud, and SaaS platforms by pulling your balance data. On the Free tier, Costbrake runs hourly checks; Pro and Team plans check every 5 minutes, scaling up to 2-minute checks during sudden spikes. A spike alert fires if an hour costs at least $5 more than usual and exceeds 3x your 7-day baseline. You can configure budgets with alerts at 50%, 80%, and 100%, alongside month-end forecasts based on the prior 72 hours.

Connecting Hyperbolic to Costbrake lets you watch your remaining credits drop, though keep in mind that a manual top-up between polling checks will offset the apparent spend for that window.

One security detail matters here: Hyperbolic API keys currently carry full account access, which means a leaked key can rent GPUs or trigger high-volume inference under your balance. Hyperbolic doesn't offer granular read-only keys, and Costbrake doesn't have brakes for this provider yet. If a key ever leaks or starts draining funds, jump straight into the Hyperbolic console under Settings > API Keys and revoke it manually. The revoked key will stop working within about a minute.

You can link your tools manually at app.costbrake.com/signin, or configure monitoring via agent setup by giving your AI coding assistant the setup URL. Check our Costbrake pricing page for tier limits and features.