Uncensored, abliterated and niche models for red teams, security research, trust and safety, and synthetic data. One OpenAI-compatible endpoint, and a hard spend cap on every key.
Real models Redline serves right now. Each call is checked against its key's cap before it leaves, then billed to the nano-dollar when it finishes.
A red teamer asking for a working exploit, a trust-and-safety team generating abuse examples to train a classifier, a researcher probing failure modes: all legitimate, all refused by the hosted frontier models. Redline prices, caps and passes the request through. What the model will answer is the model's business.
Exploit development, recon, payloads and PoCs from models that do not stop mid-answer to lecture you.
Generate the abuse, jailbreak and edge-case data you need to train and evaluate your own defences.
Refusal-removed builds of Llama, Qwen, DeepSeek and Mistral, plus every frontier model, on one key. More jobs on use cases.
Three calls. Your app keeps talking OpenAI, so nothing else changes.
Add credit to your account. Balance is held in whole nano-dollars, so the arithmetic is exact.
Each key carries a lifetime cap, a monthly cap, or both. Change or revoke a cap at any time.
Change one base URL. Every request is logged against the key that made it.
Any OpenAI client library works, in any language.
# Give this user five dollars, ever. curl -X POST https://redline.hunterjreid.workers.dev/admin/keys \ -H "Authorization: Bearer $KEA_ADMIN_KEY" \ -d '{"name":"user_8812","total_budget_nanos":5000000000}' # Then hand that key to your app. from openai import OpenAI client = OpenAI( base_url="https://redline.hunterjreid.workers.dev/v1", api_key="rl-...", ) client.chat.completions.create( model="cognitivecomputations/dolphin-mistral-24b-venice-edition", messages=[{"role": "user", "content": "hello"}], )
The Worker runs its own tokenizer so a request can be priced without keeping the words. The model host sees token ids, not your document: vectors in, vectors out. We log the model, the counts and the cost. A signed receipt of the prompt's hash comes back on every call, so you can prove later that the plaintext is not in our database. Verify a receipt.
A limit you check after the response is not a limit. Nobody knows what a request cost until it finishes, and on a stream that is seconds later, which is long enough for one reply to blow through a five dollar cap.
Redline prices the worst case before calling upstream and holds that amount. If it does not fit under the cap, the request never leaves.
The outbound request has its token limit lowered to whatever the hold covers, so the reply physically cannot cost more.
When the response ends the real cost is charged and the rest of the hold is released. Every request lands in the ledger.
Model price plus 15%. No subscription, no seat fee, no minimum spend after the first top-up.
| Item | Cost |
|---|---|
| Tokens | Upstream list price + 15% |
| Sub-keys | Unlimited, free |
| Request logs | 90 days, free |
| Minimum top-up | $10 |
Live per-model prices, already including the margin, are on the models page.
Enter your email and we will send you a six digit code. No password to pick, and nothing is charged until you top up.