StarSeaAPI
FAQ
Is the API OpenAI-compatible?
Yes. All text models use the OpenAI wire format for /v1/chat/completions and /v1/responses, including streaming, tool calls and image content blocks. Anthropic Messages and Gemini-style paths are also terminated natively.
How do I switch from OpenAI with minimal code change?
Change base_url to this endpoint and swap the API key. Nothing else in your request or response handling needs to change.
How is pricing calculated?
Per-token models are billed on input and output tokens separately, with output typically a multiple of input. Image models are billed per request. Every model page lists both the per-million-token rate and the multiplier, and the pricing table shows the full catalogue side by side.
Do cached tokens cost less?
Yes — repeated prompt prefixes are billed at a discount, because the upstream keeps the corresponding KV cache and does not need to re-run prefill. The platform routes requests so that a continuing conversation stays on the same upstream, which is what makes the discount actually apply. Put stable content (system prompts, reference documents) first and volatile content last.
What happens when an upstream provider has an incident?
Models are served through multiple independent upstream routes. Requests that hit a failing route are retried on another one, and channels are automatically taken out of rotation when their error rate crosses a threshold, then brought back when they recover.
Can I use models that do not natively support image input with images?
Yes. For text-only models the gateway reads the image and injects a description into the prompt, so the same request body works across the catalogue. Native multimodal models receive the original image unchanged.
Which payment methods are supported?
Alipay and WeChat Pay, with balance credited automatically after payment. Accounts are denominated in CNY; the pricing pages also show a USD equivalent for reference.
Is there a free tier?
No. There is no free quota to farm, which is also why the platform does not need aggressive rate limiting on paid traffic. The low-cost model tier is inexpensive enough to evaluate on real traffic.
How do I check my balance and usage without logging in?
Use the key query page: paste your API key and it returns balance and usage from a read-only, masked view, on a rate limit separate from the main API path.
Can I configure many clients at once?
Yes. A one-click import script for Windows and macOS writes the model list into the local config of supported AI coding clients, so you do not have to type model ids by hand. The script runs once and exits — it registers no background task.
Are model ids stable?
Ids follow the upstream naming and are kept stable. Version-suffixed variants (for example a dated build) are listed as separate ids so that a rolling upstream update cannot change behaviour under you.
Do you support streaming and long outputs?
Streaming is supported on all text models. Maximum output length is a property of each model and is listed per model; reasoning models count thinking tokens toward that budget.
What is the difference between native and bridged multimodal support?
Native means the upstream model itself accepts images and reads them directly — best for detail-sensitive tasks. Bridged means the platform performs the image understanding and passes text downstream — best when you want one cheap model to handle everything.
Can I use one key for every model?
Yes — a single key works across the whole catalogue, and quotas or model scope can be restricted per key when you need to hand out narrower access.