One API key for 72+ LLMs
GPT, Claude, GLM, DeepSeek, Qwen, Kimi, MiniMax, Grok and more — served from a single https://starseaapi.com/v1 endpoint that speaks the OpenAI wire format, with per-model pricing you can audit.
# one client, every provider from openai import OpenAI client = OpenAI( base_url="https://starseaapi.com/v1", api_key="sk-..." ) resp = client.chat.completions.create( model="deepseek-v4.1-flash", # or any of 72 messages=[{"role": "user", "content": "Hi"}] ) # billed at the rate printed on /pricing
Where every model actually sits
Intelligence index plotted against output price. Points on the upper-left frontier deliver the most capability per dollar — those are the models we surface first.
Why StarSeaAPI
Multi-provider supply, no single point of failure
Every model is served through more than one independent upstream route. A provider incident, a quota exhaustion or a regional failure degrades capacity rather than availability, and traffic is re-routed without intervention.
↗Cache-aware routing that lowers the real cost of long sessions
Upstream model providers bill repeated prompt prefixes at a discount, but only if the request lands on the node that already holds the cache. The router anchors a conversation to the same upstream for as long as it runs, so agent loops and long documents keep hitting cache instead of re-paying prefill on every turn.
↗Capability bridging closes the gaps between providers
Suppliers are uneven: some models are strong but text-only, some expose only one endpoint style, some speak a protocol your client does not. The gateway closes those gaps — images can be sent to text-only models, image/video/speech generation is presented through one contract, and three protocol styles are terminated natively.
↗Reliability designed in, not patched in
Health checks run on short and long windows so that spikes are not mistaken for failures and slow degradations are not missed. Deployments ship as a parallel environment validated before traffic moves, with the previous version kept as a fallback path, so releases do not interrupt in-flight requests.
↗Pricing you can audit
Every model page shows the input rate, the output multiplier and, for image models, the per-request price. Cache hits are metered and billed at the discounted rate rather than silently averaged, and per-key usage can be checked without contacting anyone.
↗Self-serve from first call to production
The model catalogue tracks upstream availability automatically, a one-click import script configures your local clients, and there is no seat licence, sales call or minimum commitment between signing up and shipping.
↗Featured models
| Models | Providers | Input $/M | Output $/M | Intelligence index | Coding index |
|---|---|---|---|---|---|
| claude-opus-5-5 frontier | Anthropic | $2.00 | $10.00 | 57.6 | — |
| claude-sonnet-5-5 frontier | Anthropic | $1.00 | $5.00 | 56.0 | — |
| claude-fable-5-1 | Anthropic | $5.00 | $25.00 | 53.4 | 81.6 |
| gpt-6-astra | OpenAI | $5.00 | $25.00 | 52.7 | 76.9 |
| gpt-6.1-sol | OpenAI | $1.00 | $5.00 | 51.8 | — |
| claude-opus-5 | Anthropic | $2.50 | $12.50 | 50.8 | 78.0 |
| claude-fable-5 | Anthropic | $5.00 | $25.00 | 49.6 | 76.5 |
| muse-spark-1.3-contributor frontier | — | $0.10 | $0.20 | 48.1 | 75.8 |
| gpt-6-sol | OpenAI | $1.00 | $5.00 | 47.6 | — |
| gpt-5.6-sol | OpenAI | $2.50 | $15.00 | 47.0 | 77.4 |
| grok-4.7 | xAI | $2.00 | $6.00 | 46.4 | — |
| mimo-v2.6-pro | Xiaomi MiMo | $0.46 | $0.92 | 46.3 | — |
Capabilities
Chat & general-purpose models
66 models
↗Reasoning models
66 models
↗Coding models
66 models
↗Vision & multimodal input models
46 models
↗Image generation models
5 models
↗Long-context models
65 models
↗Low-cost models
26 models
↗Frontier models
12 models
↗OpenAI-compatible API
67 models
↗Quickstart
Create an API key
Sign up, then create a token in the console. A key works across the whole catalogue; scope and quota can be restricted per key.
Point your client at the endpoint
Set base_url to https://starseaapi.com/v1 and use the key. Any OpenAI-compatible SDK, IDE or desktop client works.
Pick a model
Start with a low-cost tier for prototyping. Every id in the catalogue can be swapped without touching code.
Watch cost, not just quality
Output tokens are usually priced at a multiple of input, and reasoning models emit thinking tokens. Keep stable prompt prefixes first so caching applies.
Code
pip install openai
from openai import OpenAI
client = OpenAI(
base_url="https://starseaapi.com/v1",
api_key="sk-..."
)
r = client.chat.completions.create(
model="deepseek-v4.1-flash",
messages=[{"role": "user", "content": "Hello"}]
)
print(r.choices[0].message.content)