Billing, Credits & GPU Rates
Everything the API can charge you, on one page. Pricing is pure
pay-as-you-go: no subscriptions, no minimum spend. All prices are expressed
in credits (1 credit = $0.01), drawn from a single balance that
covers every model. Every response reports exactly what the call cost in its
usage object and headers, so your bill is reconstructible request by
request.
Credits
| 1 credit | $0.01 |
| Free tier | 500 credits on sign-up |
| Top-ups | Any amount, $0.01 per credit, from the dashboard |
InertialAI-0.1 — POST /v1/chat/completions
The chat model for real-world data bills per
token, metered from the usage object on every response:
| Component | Rate |
|---|---|
| Input tokens | $0.25 per 1M tokens (0.000025 credits / token) |
| Output tokens | $1.25 per 1M tokens (0.000125 credits / token) |
logprobs / top_logprobs | Free — no surcharge |
Cold starts bill like any other request — only the tokens generated, no extra charge for the load time.
Chronicle finetuning — POST /v1/finetune/jobs
Training is a $5 base fee per job plus the data you train on — a per-model rate per million training units (1 unit = 1 text token; 4 series points = 1 unit). The quote is shown before you pay and frozen at payment:
| Base model | Full finetune (listed rate) | LoRA — 75% off (default) |
|---|---|---|
inertialai-embed | $20 / 1M units | $5 / 1M units |
chronicle | $20 / 1M units | $5 / 1M units |
inertialai-forecast | $48 / 1M units | $12 / 1M units |
So a 10M-unit LoRA fine-tune of chronicle costs $5 + $50 = $55.
Improvement guarantee: if the tuned model does not beat the frozen base on the held-out eval, the fee auto-refunds. Cancelling a job while it is queued or training also refunds the fee.
Serving your models — POST /v1/endpoints/{model}/predict|embed
Every deployment runs on a dedicated GPU, billed by GPU time at that GPU's
hourly rate. A call is metered wall-clock, cold-start included;
always-warm replicas bill hourly while reserved. A scale-to-zero
endpoint costs nothing between calls. (Reported latency is separate from
billed time: each call's usage.latency_ms excludes cold start, which is
surfaced on its own as usage.cold_start_ms.)
| GPU | Price per hour |
|---|---|
| T4 | $0.97 |
| L4 | $1.32 |
| A10G | $1.82 |
| L40S | $3.22 |
| A100 40GB | $3.47 |
| A100 80GB | $4.13 |
| H100 | $6.52 |
Region pinning multiplies these rates: us-east/us-west ×1.25, eu-west
×1.5 (auto carries no multiplier). Every endpoint has a spend cap
(default $50): alert at 80%, automatic suspension at 100%. Each inference
response carries its metered cost in usage.compute_cents.
Storage — tuned weights, eval reports, validation reports — is free. See Finetuning Pricing for details and the pilot program.
Other endpoints
Embeddings — POST /v1/embeddings
| Component | Rate |
|---|---|
| Input tokens (text + time-series) | 0.000025 credits / token ($0.25 per 1M tokens) |
Output vectors are not separately charged.
Forecasts — POST /v1/forecasts
| Component | Applies to | Rate |
|---|---|---|
| Request fee | every forecast | 1 credit ($0.01) |
| Data points | every forecast | 0.015 credits / point ($0.15 per 1K); points = horizon × quantiles |
| Reasoning compute | calls that invoke reasoning (context / image_urls present) | metered per call — typically a few cents, scaling with context length |
A forecast-only call is fully deterministic before you send it:
credits = 1 + 0.015 × horizon × quantiles. See
Forecasting Pricing for worked examples and the
automatic routing rules.
Rate limits
| Surface | Default limit |
|---|---|
| Chat completions | 300 requests / minute / account |
| Finetune job operations | 60 requests / minute / account |
| Finetuned inference (predict/embed) | 300 requests / minute / account |
Exceeding a limit returns 429 Too Many Requests; retry after a short
backoff. Higher limits are available — talk to us.