Migrate from OpenAI · Cobble Docs

Cobble is a drop-in replacement: same API shape, same SDKs, open models on sustainable infrastructure. A typical migration is three changed lines and one model-name swap.

The three lines

python

client = OpenAI(
    base_url="https://api.cobble.network/v1",   # was api.openai.com/v1
    api_key=os.environ["COBBLE_API_KEY"],        # was OPENAI_API_KEY
)
# model="gpt-4o-mini"  →  model="qwen/qwen3.6-27b"

That's the entire code change. Everything else on this page is about choosing models and knowing the behavioral differences.

Model mapping

There's no exact equivalence between proprietary and open models — but there are sensible starting points by job:

If you were using	Start with	Why
GPT-4o / GPT-5.x for reasoning & agents	`qwen/qwen3.5-122b-a10b`	Flagship-class reasoning, 256K context
GPT-4o-mini / small tier for general chat	`qwen/qwen3.6-27b` or `gemma4/gemma4-26b-a4b`	Strong quality at mid-tier cost
The cheapest tier for classification, routing, drafts	`gemma4/gemma4-e2b` or `mistral/mistral-nemo-12b`	130–160 tok/s, pennies per million tokens
A coding model	`deepseek/deepseek-v4-flash`	1M context, agentic coding specialist
`text-embedding-3-small`	`nomic/nomic-embed-1.5`	$0.01/1M tokens
`text-embedding-3-large`	`qwen/qwen3-embedding-8b`	32K context window

Treat these as starting points and A/B your real prompts — open models differ more from each other than proprietary tiers do, and switching between them is a one-string change.

What behaves differently

Token counts change. Different models use different tokenizers; the same prompt may be 10–20% more or fewer tokens. If you enforce max_tokens budgets or truncate context by token count, re-measure — don't port hardcoded numbers.
Prompts tuned for one model need a pass on another. System prompts with heavy "you are ChatGPT"-era steering often over-constrain open models. Start with your prompt as-is, then simplify.
JSON mode and tool calling work, but validate. All catalog generative models support the standard tools and response_format parameters; smaller models are less reliable at complex nested schemas. If structured output matters, test at your target model size.
Unsupported OpenAI parameters are ignored, not rejected. Requests won't fail on OpenAI-specific extras, but features like the Assistants API, built-in web search, or proprietary file tooling have no equivalent here — Cobble is the inference layer; orchestration stays in your code or agent framework.
No data leaves for model training. Your prompts and outputs are not used to train anything.

Migration checklist

Create a key, set COBBLE_API_KEY, verify with the quickstart curl.
Change base URL and key in one non-critical service.
Pick the mapped model; run your existing eval or a golden-prompt set against it.
Compare cost: your plan window covers steady traffic; spiky overflow bills from the wallet at catalog rates.
Ship, watch error rates for a day, then migrate the rest.

Coming from OpenRouter or another router? Even less changes — your model IDs are likely already in org/model format; check them against /v1/models.