How Inference is metered on one credit balance
Base Rate Summary: text-embed-v1 is 3 credits per 1,000 tokens. sentiment-multilingual-v1 is 2 credits per 1,000 tokens. Vision models are 3 credits per image. Custom inference endpoints run your model and meter their compute.
Catalogue models: priced per unit of meaning
The catalogue exists so common intelligence does not start with a vendor negotiation. Embedding a support corpus, classifying tone across languages, scoring an uploaded image — each has a posted unit price in credits, invoked by name from project code. Usage is visible beside the functions that caused it, which ends the familiar archaeology of which feature spent the model budget.
Custom endpoints: your model, platform compute
When the catalogue does not carry the model — a fine-tune, a specialist classifier, a model you are contractually bound to — declare an inference endpoint with its runtime and run it as project infrastructure. The meter becomes the endpoint’s compute (0.01 credits per GiB-second while it works), and callers cannot tell a catalogue model from yours: same project, same balance, same plan review when it changes.
- No per-seat or platform fee around model calls — the unit price is the price
- Model choice is per call site: swap text-embed-v1 for a custom endpoint without touching the corpus tables
- Vision at 3 credits per image makes a 10k-upload month a ≈30,000-credit line you can state in advance
Credit Consumption Scenarios vs Hyperscaler Billing
| Workload Scenario | Monthly Resource Footprint | KERNSTACK Credits | Hyperscaler / SaaS Equivalent |
|---|---|---|---|
| Light | FAQ assistant: 2M tokens embedded once, 200k tokens/month of queries, 1k images checked | ≈6,600 credits to build, then a few hundred monthly | Embedding and chat billed by a model vendor under a separate key, card, and data-processing agreement |
| Steady | Support intelligence: 50M tokens/month embedded and classified, 20k images moderated | ≈150,000 + 40,000 + 60,000 by model line | Three vendor meters (embed, classify, vision) if bought separately, each with its own minimums |
| Heavy | Corpus programme: 500M tokens embedded for a re-index, vision on 200k images | ≈1,500,000 embedding + 600,000 vision, one-off heavy month | Rate limits and batch tiers negotiated per vendor; the project’s other meters continue separately regardless |
Frequently asked questions
Not for catalogue models — they are called by name from project code and metered to the project balance. Keys only enter the picture if you deliberately wire an external provider into a worker yourself.
Yes, as a declared inference endpoint with its runtime, metered on its compute. The calling code treats it like a catalogue model, so migrations between catalogue and custom are call-site changes, not architecture changes.
Attribution is structural: every call is made by a function, worker, or workflow step inside a project whose balance you watch. Setting a project’s scale is a budget conversation with the usage already itemised by the code that spent it.
Related capabilities, architectures, and comparisons
Ship your stack on the Software Lifecycle Cloud.
Closed preview on dedicated European infrastructure. Bring a repository or start from a blueprint.