SERVE / PRIVATE ENDPOINTS

Private endpoints your SDK already speaks

Swap the base URL in your OpenAI or Anthropic client. The rest of your code stays the same, and every model you train serves behind its own endpoint, with keys, usage metering and scaling behind it.

PythonTypeScriptcURL
from openai import OpenAI

client = OpenAI(
    base_url="https://api.nucleus-ai.io/v1",   # the one line that changes
    api_key=NUCLEUS_API_KEY,
)

client.chat.completions.create(
    model="nucleus://run_3Bz:train:0/sampler_weights/000240",
    messages=[{"role": "user", "content": "Triage this ticket"}],
)
200 OK
served
model
…/sampler_weights/000240
prompt_tokens
512
completion_tokens
128
latency
142 tok/s

One line changes.

Point your OpenAI or Anthropic client at a Nucleus base URL and keep the rest of your code.

Your own endpoint.

Every model you train serves behind its own endpoint, isolated to your account.

Scales with traffic.

Serving scales with your usage, and Nucleus runs the servers.

Every token accounted for.

The usage API reports served tokens for any date range, grouped by day, project or model.

How it works

Two routes, one key, everything behind it

Point any OpenAI client at https://api.nucleus-ai.io/v1 and authenticate with your Nucleus key. The model parameter takes a base model ID or the nucleus:// path of your fine-tuned weights, so the same client serves both. Chat completions renders your messages with the model's chat template on the server, and stream: true returns tokens as they are generated.

OpenAI-compatible API reference
Results

Results from LLMs powered by Nucleus

$73.4K

monthly revenue misattributed, found

47

trucks rerouted while they were still moving

$177K

downtime prevented this month

Same day

from site reading to client report

STORIES

Three workflows served this way

Field engineering

Cable survey readings processed while the crew was still on site.

Same dayfrom site reading to client report
9systems on one thread
IT operations

Every incident past its four-hour SLA, escalated by name.

8incidents past their four-hour SLA
6.2 hrsaverage time to resolution
Fleet operations

47 trucks rerouted while they were still moving.

$2.4Kdaily fuel saved across the fleet
23delays resolved without a call

Private Endpoints at a glance

Routes

  • Chat completions, POST /v1/chat/completions
  • Completions, POST /v1/completions
  • Messages, POST /anthropic/v1/messages
  • Count tokens, POST /anthropic/v1/messages/count_tokens
  • Streaming over text/event-stream on every route
  • Reasoning controls: reasoning_effort, separate_reasoning; thinking, output_config.effort
  • Prompt caching and citations are unsupported on the Anthropic route

Models

  • Any base model from the catalogue: 25 open models across language, vision, image, video and audio
  • Your fine-tuned checkpoints, by nucleus:// path
  • A fine-tuned model serves at its base model's per-token rate
  • Base model IDs and paths in the same model parameter

Keys and access

  • Scopes read, train, infer, admin
  • Project binding, so a key sees one project's resources
  • nk_live_ production and nk_test_ sandbox prefixes
  • Rotation with two keys valid at once
  • GET /v1/auth/verify to see what a key can do
  • Rate limits per key, reported in X-RateLimit headers

Operations

  • Usage by day, project or model, GET /v1/usage
  • Job events over SSE, resumable with after
  • Audit log by UTC day, GET /v1/audit-log
  • Automatic scaling with usage
  • SSL, firewall and DDoS protection built in
  • Serverless, dedicated or reserved capacity
PRICING

Served tokens

Served tokens

From $0.15per 1M tokens

Calls to your private endpoints are metered per million input and output tokens, through the OpenAI- and Anthropic-compatible APIs. Input and output tokens cost the same. A fine-tuned model is served at the same per-token rate as its base model. There is no premium and no standing per-model fee on serverless.

Pick your path

The three ways to run an endpoint

Serverless

Pay per token, idle costs nothing

Explore Serverless

Dedicated GPUs

L40S to B200, by the second

Explore Dedicated GPUs

Reserved capacity

Guaranteed terms, one to six months

Explore Reserved capacity

Compatible APIs

Docs

One key, two SDK families

Explore Compatible APIs

Ready to change one line?

Point your OpenAI or Anthropic client at a Nucleus endpoint today, or talk to sales about dedicated and reserved capacity.