API documentation

Arkheon mirrors the OpenAI API. Change base_url on https://arkheonchat.com/v1 and the key — the code keeps working, now with every model and a ruble account.

Quick start

1. Sign up and confirm your email.
2. Top up on the page Pricing.
3. Create a key in your account. A key is shown once.

Copyfrom openai import OpenAI

client = OpenAI(base_url="https://arkheonchat.com/v1", api_key="ark_...")

resp = client.chat.completions.create(
    model="anthropic/claude-sonnet-4",
    messages=[{"role": "user", "content": "Check this contract for risks"}],
)
print(resp.choices[0].message.content)

Keys

Pass the key in the header Authorization: Bearer ark_…. Up to 10 active keys per account, each can be revoked in your account. A key opens the API only — not the account and not the password.

POST /v1/chat/completions

Supported: model, messages (text, image_url, file), max_tokens / max_completion_tokens, temperature, top_p, stop, stream, tools, tool_choice, parallel_tool_calls, response_format, seed, presence_penalty, frequency_penalty, reasoning, modalities. Other fields are ignored; n must be 1.

The response is the standard OpenAI object. In addition it carries arkheon.credits_charged and arkheon.balance. If max_tokens is not given, 2048 is used; the maximum is 16,384.

Web search is switched on with the suffix :online of the model, for example openai/gpt-4o-mini:online.

Model auto picks the model for the request by itself: code, calculation, analysis of long documents, text or a quick answer. If "ask before expensive models" is on in your account, the API takes a model below the threshold instead of asking; pass "arkheon": {"allow_expensive": true} in the request body to allow an expensive one. The actual model comes back in arkheon.model, the reason for the choice is in arkheon.auto.

Streaming

"stream": true — the answer comes as text/event-stream, the last chunk contains usage, then data: [DONE]. Works with stream=True in the SDK.

Voice: transcription and speech

The same calls as OpenAI's. The names whisper-1, gpt-4o-transcribe, tts-1, gpt-4o-mini-tts and OpenAI's voices (alloy, coral, verse and the rest) are accepted and lead to our default models.

from openai import OpenAI
client = OpenAI(base_url="https://arkheonchat.com/v1", api_key="ark_...")

text = client.audio.transcriptions.create(
    model="whisper-1", file=open("voice.ogg", "rb")).text

with client.audio.speech.with_streaming_response.create(
        model="tts-1", voice="coral", input="Добрый день!") as r:
    r.stream_to_file("answer.mp3")

POST /v1/audio/transcriptions

A file up to 25 MB and 60 minutes: mp3, wav, flac, ogg, m4a, webm, mp4. The answer comes as json, text, verbose_json, srt or vtt. You pay for the length of the recording; a recording with no speech in it is free.

POST /v1/audio/speech

Up to 4096 characters a request. Formats: mp3, opus, aac, flac, wav and pcm (24 kHz, 16-bit, mono); the audio streams as it is made. The price is counted by characters and comes in the header x-arkheon-credits-charged. The text goes through the same check as chat messages.

Models, voices and prices are in the answer of GET /v1/audio/models.

Live voice: /v1/realtime

A WebSocket in the OpenAI Realtime protocol. Clients written for gpt-realtime (the OpenAI SDK, Pipecat, LiveKit Agents) connect by changing the address and the key. Inside: speech recognition, a chat model from the catalog and speech sentence by sentence; the answer starts playing about 2–3 seconds after the end of a phrase.

wss://arkheonchat.com/v1/realtime?model=google/gemini-2.5-flash-lite

import OpenAI from "openai";
import { OpenAIRealtimeWS } from "openai/realtime/ws";

const client = new OpenAI({ baseURL: "https://arkheonchat.com/v1", apiKey: "ark_..." });
const rt = await OpenAIRealtimeWS.create(client, { model: "gpt-realtime" });
rt.on("response.output_audio.delta", (e) => play(Buffer.from(e.delta, "base64")));
rt.send({ type: "input_audio_buffer.append", audio: pcm16.toString("base64") });

• The model is set with the parameter model: any chat model from the catalog. Without it, or with an OpenAI model name, the answer comes from google/gemini-2.5-flash-lite.
• Audio: pcm16 at 24 kHz, or G.711 (audio/pcmu, audio/pcma) for telephony.
• Supported: server-side pause detection or manual commits, interrupting, cancelling an answer, truncating what was heard, function calling, text-only answers, GA and beta events.
• Not supported: WebRTC, SIP, images and MCP.

From a server, the key goes in the header Authorization. Don't give the key to a browser: your server gets a one-time token with POST /v1/realtime/client_secrets, and the page passes it in the subprotocol openai-insecure-api-key.<token>. The token lives 60 seconds and opens one connection; an ark_ key in the subprotocol is refused.

You pay per answer, for recognition, the model and speech together, usually 1 credit an answer. An answer cancelled before its first sound is free. Up to 5 sessions at a time, each up to 30 minutes. Audio and transcripts are not stored.

GET /v1/models

The list of available models. credits_per_answer — the approximate price of a typical answer (300 tokens in and 500 out), not a flat price: the actual cost by tokens is charged, and prices per 1000 tokens are in credits_per_1k_input_tokens and credits_per_1k_output_tokens. Long context — documents, notes, history — is paid for again in every request. Identifiers of the form vendor/model. Below are the cheapest of 178.

The full list is in the response /v1/models or in the model picker in the chat.

Billing

Before a request an estimate is held on the balance (it accounts for max_tokens). After the answer the actual cost reported by the provider is charged and the rest of the hold is returned at once. 1 credit ≈ 1 ₽, the minimum charge is 1 credit. Empty answers and provider errors are not charged. Balance: GET /v1/balance.

GET /v1/usage

Statistics for the key the request is signed with: ?days=30 (1–90). The answer contains total (requests, tokens, credits), by_day (a dense series by day), by_model. A key sees only itself; the whole picture of the account is in your account.

curl https://arkheonchat.com/v1/usage?days=7 -H "Authorization: Bearer ark_..."

{"object":"usage","scope":"key","key":{"id":12,"name":"Бот поддержки","prefix":"ark_3f"},
 "period":{"from":"2026-08-27","to":"2026-09-02","days":7},
 "total":{"requests":418,"prompt_tokens":91230,"completion_tokens":40211,"cached_tokens":0,"credits":612},
 "by_day":[{"date":"2026-08-27","requests":51,"credits":80,...}],
 "by_model":[{"model":"openai/gpt-4o-mini","requests":390,"credits":402,...}]}

Errors and limits

Errors have the OpenAI shape: {"error": {"message", "type", "code"}}.