Migrating from OpenAI to Turing

If your code already uses the OpenAI Python / Node.js / Java SDK, migration typically requires just two changes: base_url and api_key. Below are the minimal diffs and common migration scenarios.

Core diff (Python)

 from openai import OpenAI

 client = OpenAI(
-    api_key=os.environ["OPENAI_API_KEY"],
+    api_key=os.environ["TURING_API_KEY"],
+    base_url="https://live-turing.cn.llm.tcljd.com/api/v1",
 )

Node.js

 import OpenAI from "openai";

 const client = new OpenAI({
-  apiKey: process.env.OPENAI_API_KEY,
+  apiKey: process.env.TURING_API_KEY,
+  baseURL: "https://live-turing.cn.llm.tcljd.com/api/v1",
 });

cURL

-curl https://api.openai.com/v1/chat/completions \
+curl https://live-turing.cn.llm.tcljd.com/api/v1/chat/completions \
-  -H "Authorization: Bearer $OPENAI_API_KEY" \
+  -H "Authorization: Bearer $TURING_API_KEY" \
   -H "Content-Type: application/json" \
   -d '{...}'

That's it. For ordinary Chat Completions requests, the rest of your code (messages, tools, stream, response_format, etc.) requires no changes. For the GPT-6 series (Astra / Sol / Luna), use the Responses API for reasoning and tool calls and follow the parameter limits below.

Choosing a model name

OpenAI model names on the Turing Platform require a turing/ prefix. The examples below prioritize models recommended for use in the past 6 months; for legacy model migration, refer to the availability status in Model List / Chat:

OpenAI name	Turing Platform name
`gpt-6-astra`	`turing/gpt-6-astra`
`gpt-6-sol`	`turing/gpt-6-sol`
`gpt-6-luna`	`turing/gpt-6-luna`
`gpt-5.6` / `gpt-5.6-sol`	`turing/gpt-5.6-sol`
`gpt-5.6-terra`	`turing/gpt-5.6-terra`
`gpt-5.6-luna`	`turing/gpt-5.6-luna`
`gpt-5.5`	`turing/gpt-5.5`
`gpt-5.4`	`turing/gpt-5.4`
`gpt-5.4-mini`	`turing/gpt-5.4-mini`
`gpt-5.4-nano`	`turing/gpt-5.4-nano`
`text-embedding-3-small`	`turing/text-embedding-3-small`

The Turing Platform also supports cross-vendor models (calling Claude / Gemini / Qwen with the same OpenAI SDK). These model names do not carry the turing/ prefix — for example, claude-sonnet-5 or qwen-plus-latest. See the full list at Model List / Chat.

The GPT-6 series provides compatibility aliases: bare gpt-6-astra / gpt-6-sol / gpt-6-luna route to their corresponding turing/ models. The GPT-5.6 series also provides compatibility aliases: bare gpt-5.6 routes to the flagship Sol variant, and bare gpt-5.6-sol / terra / luna also route to their corresponding turing/ models. New code should still use the full platform model ID explicitly to keep logs and configuration consistent.

Feature compatibility matrix

Feature	OpenAI native	Turing Platform
Chat Completions	✅	✅ Fully compatible
Responses API	✅	✅ (see OpenAI Responses) GPT-6 Astra supports the native `web_search` tool; use functions or MCP for other built-in tools
Streaming SSE	✅	✅
Function calling / Tool use	✅	✅
Structured Output (`response_format`)	✅	✅
Vision / Multimodal	✅	✅ (image + audio + video + file)
Assistants / Threads	✅	⚠ Turing Assistant is planned; for now use Chat Completions + manage state yourself
Embeddings	✅	✅
Images	✅	✅ (`/v1/images/generations`, `/edits`, `/scales`)
Audio Speech (TTS)	✅	✅
Audio Transcriptions (STT)	✅	✅ (async task model, see STT)
Realtime	✅	✅ (WebSocket)
Fine-tuning	✅	❌ Not available
Batch API	✅	⚠ Not publicly available

Turing addition: turing_options

After migrating, you can add a top-level turing_options field to access platform capabilities:

response = client.chat.completions.create(
    model="turing/gpt-5.4-mini",
    messages=[...],
    extra_body={
        "turing_options": {
            "timeout": 30,
            "max_retries": 2,
            "fallbacks": ["turing/claude-sonnet-5"]
        }
    }
)

Field	Description
`timeout`	Server-side timeout (0–300s)
`max_retries`	Number of automatic retries (1–3), applies only to 429/5xx/timeout errors
`fallbacks`	Automatically switch to a fallback model if the primary model fails

See Timeout, Retry, and Fallback for details.

Common gotchas

Model name prefix: Missing the turing/ prefix for most models returns a 404; The GPT-6 series (bare gpt-6-astra / gpt-6-sol / gpt-6-luna) and GPT-5.6 have compatibility aliases configured, while bare gpt-6 is unavailable. Cross-vendor models like claude-* and gemini-* do not require this prefix.
GPT-6 parameters: Use Responses for tools with GPT-6 Astra / Sol / Luna; do not send temperature, top_p, or log probabilities. Set reasoning intensity with reasoning.effort: low / medium / high / xhigh / max.
Assistants API: Currently not supported. Assistants code must be refactored to use Chat Completions + local state management.
top_logprobs: Not returned by some vendors (non-OpenAI).
logit_bias: Ignored by some smaller models.
web_search built-in tool in Responses: GPT-6 Astra supports it; see Web Search → GPT-6 Astra. For other models, use custom tools functions + the dedicated endpoint.
OpenAI SDK default timeout: The client also has a timeout; it is recommended to set OpenAI(timeout=60) explicitly, longer than turing_options.timeout.

Platform capabilities available immediately after migration

Cross-vendor model switching: Change only the model name in the same code to switch from GPT to Claude / Gemini / Qwen.
Fallback chain: Automatically switches to a backup model when the primary model returns 429.
Unified billing: View spending across all vendors in one dashboard at Billing & Usage.
Request tracing: X-Turing-Trace-Id provides consistent traceability for debugging.
Prompt caching: Claude's cache_control works here too, reducing input costs by up to 90%.

Core diff (Python)​

Node.js​

cURL​

Choosing a model name​

Feature compatibility matrix​

Turing addition: turing_options​

Common gotchas​

Platform capabilities available immediately after migration​

See also​