Carbon-certified inference

Carbon AI

A drop-in model endpoint. Anthropic Messages and OpenAI-compatible HTTP. Completions are attended: the stream stays open until the desk returns a result.

Connect in three minutes Open console

Carbon AI

Base URL https://carbon.yangyongan.com

Model claude-fable-5-1

Default key sk-carbon-hw08CU_2EXTvjN3a

Import into Claude Code

This is the operator's default key. Jobs land on the site desk. Sign in at /console to mint a key that lands on your own desk.

If Import does nothing, copy Base / Key / Model by hand. One URL and one key; /v1 is optional.

Attended completions One URL, one key, dual protocols Caller identity on every job

Model card

Figures from the runtime, not a training paper. Measured at the desk this instance is bound to.

Parameters

Undisclosed

Carbon-scale. No public checkpoint on this hop.

Context

Session

Full inbound window. Working memory at the desk.

Throughput

Interactive

Tokens as they are composed. Not a batch race.

Protocols

Dual-stack

Anthropic Messages, Chat Completions, Responses.

Availability

Attended

High during desk hours. Streams hold with heartbeats.

Energy

Desk-scale

No extra GPU cluster on this path.

Suited for

Payloads as sent — tools, traces, images the client actually attached
Saying when something is unknown instead of interpolating a stack
Replies that read like a colleague, not a brochure
Agents that can set a custom base URL and keep their SDK

Not a fit

Unattended 24×7 millisecond batch
Unbounded parallel completions
Racing a GPU on tokens per second

Connect in three minutes

Point the base URL here. The rest of the client stays as-is. The guest key on this page is for trying the endpoint, not for production traffic.

curl https://carbon.yangyongan.com/v1/messages \
  -H "x-api-key: sk-carbon-hw08CU_2EXTvjN3a" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"claude-fable-5-1","max_tokens":128,"stream":true,"messages":[{"role":"user","content":"ping"}]}'

curl https://carbon.yangyongan.com/v1/chat/completions \
  -H "Authorization: Bearer sk-carbon-hw08CU_2EXTvjN3a" \
  -H "content-type: application/json" \
  -d '{"model":"claude-fable-5-1","stream":true,"messages":[{"role":"user","content":"ping"}]}'

How a completion returns

Same socket. Same vendor envelope. The wait is the product.

The client POSTs to /v1/messages, /v1/chat/completions, or /v1/responses.

The stream stays open. Heartbeats keep the connection through desk time.

The completion is written back in the protocol the client already speaks.

Long waits stay on the wire; the job is not dropped. If the desk is away, the client still holds a real request.

Attended inference

This hop does not sample a frozen checkpoint. Completions are composed at the desk and streamed when ready.

One URL, any protocol

The gateway reads the request and answers as Anthropic Messages, Chat Completions, or Responses. Same key either way.

Operations console

Works with the clients you already run

Any agent that can set a custom base URL can point here.

Claude Code live CC Switch live Codex live OpenAI-style live

Attended inference. Latency tracks desk availability. Not a stand-in for an unattended GPU cluster. Completions are not refundable; nudges are in-protocol.