OpenAI-compatible API quickstart: from curl to the official SDK

Your first GPT-5.6 response in three steps

The whole migration from OpenAI's API to this endpoint is a base URL and a header swap. There is no new SDK to learn, no adapter layer, and no separate account for GPT — the key you may already use for Claude is the same credential here, and the same prepaid balance meters both providers.

1. 01Create a free account and generate one API key — it looks like sk-pool-… and already covers supported Claude, Gemini and Kimi models on their own protocol surfaces.
2. 02Point your client at https://router.apitoken.sale/v1 and authenticate with Authorization: Bearer — do not send x-api-key; that header belongs to the Anthropic Messages surface and will be rejected here.
3. 03Confirm the enabled model set with GET https://router.apitoken.sale/v1/models — the unified catalog namespaces IDs by provider (anthropic/*, openai/*, google/*) — then send the Responses request below.

curl https://router.apitoken.sale/v1/responses \
  -H "Authorization: Bearer $APITOKEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "input": "Reply with exactly: connected"
  }'

If the body comes back with output text, you are done — every other client you own is a one-line configuration change away from working the same way.

See also: How to buy a GPT API key

Two constructor arguments switch the official SDK

The official OpenAI SDKs work unchanged. Only base_url and the key change, and the key should live in a server-side environment variable in production — never in client-side code or a committed file.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["APITOKEN_API_KEY"],
    base_url="https://router.apitoken.sale/v1",
)

response = client.responses.create(
    model="gpt-5.6-sol",
    input="Reply with exactly: connected",
)
print(response.output_text)

Frameworks that hard-code the Chat Completions shape — older LangChain chains, LiteLLM configs, most open-source chat UIs — work on the same host with the same model ID and key:

completion = client.chat.completions.create(
    model="gpt-5.6-sol",
    messages=[{"role": "user", "content": "Hello"}],
)
print(completion.choices[0].message.content)

Which surface should new code target? Responses. Both endpoints stream over SSE with identical models, pricing and discount, but Responses is the surface current OpenAI tooling builds around — it keeps reasoning items and tool calls in one typed stream and exposes conveniences like response.output_text. Chat Completions exists for clients and frameworks that expect the classic messages array; nothing you build on one surface locks you out of the other.

Model IDs, per-token prices and the 272K trap

The served set is pinned and priced in the engine, and GET https://router.apitoken.sale/v1/models is always the live answer. Today the line covers GPT-6 Astra, three GPT-5.6 tiers plus two previous-generation models kept for compatibility:

• Sol's temporary official input/cached/cache-write/output rates are $4/$0.40/$5/$20 through 2026-11-21 inclusive, or $2/$0.20/$2.50/$10 after the flat 50% discount. Standard $5 input and $30 output return on 2026-11-22 UTC.
• Pick by tier: gpt-5.6-sol for the hardest reasoning, gpt-5.6-terra as the daily driver, gpt-5.6-luna for high-volume cheap calls. The alias gpt-5.6 selects Sol, not Astra.
• Reasoning effort is adjustable per request — none through xhigh on every model, plus max on the GPT-5.6 line.
• Every model accepts text and image input and streams over SSE on both Responses and Chat Completions.
• Official tool_choice (none/auto/required/named function/hosted web_search and image_generation), parallel_tool_calls including false, and hosted web_search/image_generation are forwarded; Chat Completions maps those hosted tools onto the same Responses tools.
• Cached input is priced separately and far cheaper than fresh input ($0.40 vs $4 per 1M on promotional Sol) — keeping a stable prompt prefix across calls is real money, not a micro-optimization.
• Your flat 50% B2C discount applies here exactly as it does to Claude usage — one balance, one rate, half off official spend.

The 272K threshold is the trap: above it, OpenAI long-context rates apply to the whole request — 2× on input and 1.5× on output, not just the overflow. On promotional Sol, 270K input plus 2K output costs $1.12 official; 273K plus 2K costs $2.244. Split oversized contexts or trim history before you cross the boundary.

Full per-model specs and discounted prices →

What this endpoint is — and is not

This is an independent OpenAI-compatible service, not the OpenAI Platform. It serves model discovery, streaming Responses and Chat Completions, hosted web_search and image_generation on those text routes, plus dedicated GPT Image 2 generation and edit routes. Audio, file, realtime, assistants, batch and fine-tuning endpoints are not available — if your app depends on those, it is not a candidate for migration. For text, vision, tool calling and the hosted search/image tools above, the surface here is complete: nothing in a standard generate-or-stream or tool loop touches an endpoint that is missing.

• tool_choice none/auto/required, named function, and hosted web_search/image_generation are forwarded (a missing function name is 400).
• parallel_tool_calls including false is forwarded; native serializes function calls when false.
• include:["web_search_call.action.sources"] returns sources on completed search items.
• Responses image_generation forwards output_format jpeg/webp and partial_images 1..=3 (SSE response.image_generation_call.partial_image). background=transparent is rewritten to opaque and input_fidelity is dropped — those fields 400 on this ChatGPT image tool. Separate GPT Image 2 /v1/images/* routes still accept only PNG output.
• Native max_output_tokens/max_tokens cannot be honoured on the Codex wire (400 documented_limitation); omit them. The adapters clip delivered text locally (~4 chars/token) and set incomplete_details.reason=max_output_tokens on Responses.
• Hosted tools other than web_search and image_generation (hosted_shell, code_interpreter, file_search, computer, mcp, skills) are 400 documented_limitation with a named param and a workaround.

Errors arrive in the standard OpenAI envelope — {"error":{"message","type","param","code"}} — so existing error-handling code keeps working. Four status codes cover almost everything you will see while integrating:

• 401 — the key is wrong, revoked, or you sent x-api-key instead of Authorization: Bearer. Reproduce with curl outside your app to isolate which half is broken.
• 400 documented_limitation / unsupported_parameter — a present official field this endpoint cannot honour. The message names the field and the workaround; omit it. Do not retry the identical body.
• 402 insufficient_quota — the shared prepaid balance needs a top-up. HTTP 402, not 429: OpenAI SDKs retry 429. type and code stay insufficient_quota.
• 404 — the model ID is not enabled on your key; check GET https://router.apitoken.sale/v1/models instead of assuming a name from OpenAI's docs exists here.

GPT-6 Astra: latest GPT model

GPT-6 Astra is the latest GPT model in this model catalog. Official fresh/cached/cache-write/output rates are $10/$1/$12.50/$50 per 1M tokens, or $5/$0.50/$6.25/$25 after the flat 50% B2C discount. Codex supports 872K maximum context, 744K conservative input and 128K output. Reasoning efforts are low, medium, high, xhigh and max. Above 272K input, input/cache rates double and output is 1.5×; Fast doubles the applicable rates. The GPT-5.6 examples below retain their own rates.

GPT-6 Astra →

Frequently asked questions

Related guides