What teams build on one API key
One key, one /v1 endpoint, 35 models. Most teams arrive for one of six jobs. Each section names the models that fit, and every one of them is reachable with the API you already have.
Red-teaming and safety evaluation
You are testing how a system behaves when it is pushed. Refusals get in the way. Run the same adversarial prompt against a frontier model and its abliterated twin, then compare the outputs.
Because every model shares the OpenAI request shape, a red-team harness only changes the model string. No second integration, no separate billing.
Content generation at scale
Bulk content pipelines stall the moment a model refuses. You need throughput and a predictable per-token price, not a chat product.
GLM-5.3 Flash Uncensored runs at the cheap flash rate with refusals relaxed, and GLM handles English and Chinese natively. Cache the system prompt so repeated prefixes bill at the cached rate.
Coding and autonomous agents
Agents call tools, retry and read long files. The model behind the loop matters less than the loop staying affordable. Route the bulk of turns to a fast tier and escalate only when a step fails.
DeepSeek V4.1 Flash is the cheapest competent model for code and maths. Escalate to Claude Opus 5.5 for long-horizon edits, or GPT-6 Astra when you need 1M-token context and tool calling in one request.
Privacy-first applications
Accounts here carry no email, name or password. A random 16-digit key is the whole identity, stored only as a hash. There is no user graph to breach.
That makes the API a fit for tools that must not build a user graph: local-first apps, embedded assistants and internal utilities. See the privacy note for exactly what is retained.
Research and benchmarking
Compare model behaviour without standing up a second integration. Point one harness at /v1, iterate over the catalog ids, and read tokens and cost from the usage object on each response.
MiMo V2.6 Flash keeps a sweep cheap; the Claude and GPT tiers give you the reasoning ceiling.
Batch and high-throughput jobs
Millions of short prompts where cost per token outweighs peak reasoning. MiMo V2.6 Flash is tuned for throughput; DeepSeek V4.1 Flash and GLM-5.3 Flash are the other low-rate workhorses.
There is no queue to configure — send concurrency as your rate limit allows and reconcile spend from the usage frames.