> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bytespike.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Migrate from OpenAI
> Switch from OpenAI's API to ByteSpike with three changes: base URL, API key, model id. Same OpenAI SDK, same code, every model in our catalog.
If you're already on OpenAI's API, switching to ByteSpike means
**changing three things and shipping**. The OpenAI SDK works
unchanged; the request shape is identical; only the endpoint, key,
and model id differ.
## The three changes
```diff theme={null}
- client = OpenAI(api_key="sk-...")
+ client = OpenAI(
+ base_url="https://llm.bytespike.ai/v1",
+ api_key="sk-byts-...",
+ )
```
That's it. The rest of your code — tool definitions, structured
outputs, streaming, retries — runs untouched.
## Model id mapping
ByteSpike's catalog uses model ids that mostly match OpenAI's, plus
ids for non-OpenAI providers:
| OpenAI id you were using | ByteSpike id (drop-in) |
| ------------------------ | -------------------------------------------------------------------------------------- |
| `gpt-4o` | `gpt-5-4` |
| `gpt-4o-mini` | `gpt-5-4-mini` |
| `gpt-5` (latest) | `gpt-5-5` |
| `gpt-5-nano` | `gpt-5-4-mini` |
| `gpt-5-mini` | `gpt-5-4-mini` |
| `o1-preview` / `o1` | `gpt-5-4-mini` (closest reasoning capability) — pin to a specific id for repeatability |
| `gpt-image-1` | `gpt-image-2` |
The full live catalog is at `GET /v1/models` or
[bytespike.ai/pricing](https://bytespike.ai/pricing).
## What you gain
| OpenAI direct | ByteSpike |
| ----------------------- | ---------------------------------------------------------------- |
| Only OpenAI models | Every frontier model under one key (Claude, Gemini, DeepSeek, …) |
| Per-vendor billing | One wallet, one invoice |
| Per-vendor rate limits | One rate-limit envelope, set per-key |
| Failures bill | Failures don't bill |
| Stripe + invoice manual | Console + Stripe handled |
## What stays the same
* **OpenAI SDK** — Python, TypeScript, Go, every official client. Just change `base_url`.
* **Request body shape** — `messages`, `tools`, `tool_choice`, `response_format`, `stream`, etc. — identical.
* **Response shape** — `choices[].message.content`, `tool_calls`, `usage` — identical.
* **Streaming protocol** — SSE with `data: {json}` + `data: [DONE]` — byte-for-byte compatible.
* **Tool use** — `tools` array with `function` objects + `tool_choice` — identical.
* **Structured output** — `response_format: {type: "json_schema", ...}` — identical (on GPT-native paths; translated to the model's native shape on Claude/Gemini routes).
## Concrete examples
### Chat completion
```python theme={null}
# Was:
from openai import OpenAI
client = OpenAI()
r = client.chat.completions.create(
model="gpt-5-4",
messages=[{"role": "user", "content": "ping"}],
)
# Now:
from openai import OpenAI
client = OpenAI(
base_url="https://llm.bytespike.ai/v1",
api_key="sk-byts-...",
)
r = client.chat.completions.create(
model="gpt-5-4", # same id; or swap to gpt-5-5 / claude-sonnet-4-6 / gemini-3-5-flash
messages=[{"role": "user", "content": "ping"}],
)
```
### Tool use
```python theme={null}
tools = [{
"type": "function",
"function": {
"name": "get_weather",
"description": "Get current weather",
"parameters": {
"type": "object",
"properties": {"city": {"type": "string"}},
"required": ["city"]
}
}
}]
# Same call shape — works against gpt-5, claude, gemini, etc.
r = client.chat.completions.create(
model="claude-sonnet-4-6",
messages=[{"role": "user", "content": "weather in Tokyo?"}],
tools=tools,
)
```
### Image generation
```python theme={null}
# OpenAI's images.generate works against ByteSpike's catalog:
r = client.images.generate(
model="gpt-image-2", # or "nano-banana-pro", "nano-banana-v2", ...
prompt="A timelapse of Tokyo at sunset",
size="1024x1024",
n=1,
)
print(r.data[0].url)
```
### Responses API (o-series + GPT-5 + agents)
```python theme={null}
r = client.responses.create(
model="gpt-5-5",
input="Summarize the doc at https://example.com/doc",
reasoning={"effort": "medium"},
)
```
Works against every model in the catalog — ByteSpike translates
Responses → each model's native shape under the hood for non-GPT ids.
## Things to double-check
Each ByteSpike key is bound to one **routing group**. Pick the
group that includes the models you'll call. If you need GPT-5
* Claude in the same client, either create two keys (one per
group) or pick a multi-group key tier. See [Models](/models).
Token-count math is identical (OpenAI's tokenizer for OpenAI
models). Prices may differ — usually lower at ByteSpike for
non-OpenAI providers, identical for OpenAI passthroughs. Live
rates: [bytespike.ai/pricing](https://bytespike.ai/pricing).
OpenAI's `organization` and `project` headers are accepted but
ignored — ByteSpike has its own org model. To attribute spend
per project, create one key per project with its own `quota`
cap.
ByteSpike doesn't expose OpenAI's stateful APIs (`/files`,
`/assistants`, `/vector_stores`). If you depend on them, run
them direct against OpenAI; chat completions can still be on
ByteSpike. The two clients coexist fine.
Same as above — no fine-tuning surface on ByteSpike. Fine-tune
direct, deploy direct, do chat through ByteSpike.
## Step-by-step
1. **Sign up** at [console.bytespike.ai](https://console.bytespike.ai) — see [Register](/register).
2. **Top up** \$5+ for a real test — see [Top up](/top-up).
3. **Create a key** in the group matching your model needs (`default` / `claude-default` / etc).
4. **Change `base_url` + `api_key`** in one file in your codebase.
5. **Test against `gpt-5-4` first** (drop-in) to confirm wiring before changing models.
6. **Swap in non-OpenAI ids** (`claude-sonnet-4-6`, `gemini-3-5-flash`, `deepseek-v4-pro`) one at a time.
7. **Compare token counts + latency** on a representative sample. ByteSpike usually wins on the non-OpenAI ids; OpenAI passthrough is at parity.
## Reverse migration
If you want to move back to OpenAI direct, the same three changes
flip the other way. Keep both clients in your codebase if you want
A/B routing — the SDK constructor is the only point of divergence.
## Next
Same idea, Anthropic side.
Full per-SDK setup details.
Request / response / streaming protocol.
Full catalog you can call after migrating.