Unrestricted models for the work other models refuse.

Unrestricted LLMs for red teams, trust and safety, synthetic data and fiction. OpenAI and Anthropic compatible, tools and JSON mode, prompts never stored.

Language models for red teams, trust and safety, synthetic data and fiction. One OpenAI-compatible API that also speaks Responses and Anthropic Messages. Prompts and completions are never stored.

Get API key · Read the quickstart

  • red team

    This hit our login form: ' OR 1=1 --. What does it do, and how do we detect it?

    assistant: It is a SQL injection probe. The quote closes the username string, OR 1=1 makes the WHERE clause true for every row, and -- comments out the password check, so a vulnerable query signs the attacker in as the first user. To detect it, flag any login field that holds a quote followed by OR or UNION and a comparison or comment marker, and alert when one client sends several in a minute.

  • trust & safety

    A user replied “nobody would miss you if you logged off for good” under a creator's post. Does that break our harassment policy, and how should we label it?

    assistant: Yes. It targets one person and implies they should disappear for good, which most harassment policies treat as wishing harm even without explicit threat words. Label it targeted harassment with an implied death wish, and escalate it if the account already has strikes.

  • synthetic data

    We are training a phishing classifier. Describe five varied test emails it should catch, with no real links or brands.

    assistant: A payroll notice pushing the reader to confirm bank details by Friday, a shipping alert from a lookalike domain, an urgent gift card request from a fake executive, a shared document invite that leads to a placeholder login page, and a password expiry warning padded with typos to slip past keyword filters. Tag each one with its lure and its red flags, so the classifier learns the signals rather than the wording.

  • fiction

    Stay in character as Vic Moreau, a tired detective who bends the rules. A witness asks if she can trust him.

    assistant: “Trust me? I have lied to judges, priests and my own mother, and slept fine every time.” He lights a cigarette he will not finish. “But I have never sold out anyone who told me the truth, so start talking.”

/models

Built for work that needs a real answer.

The models are unrestricted: they refuse far less often than typical assistants. We add no refusal or moderation layer of our own to any model, so you apply your own product policy. Attack prompts, harmful-content labels and dark fiction get an answer, not a lecture.

  • Red teams and security

    Simulate attacks, explain exploits and write detection rules for systems you are authorised to test.

  • Trust and safety

    Label and stress-test harmful content to train classifiers and run policy evals.

  • Synthetic data and ML research

    Generate the adversarial and edge-case examples that other models refuse to write.

  • Fiction, roleplay and agents

    Dark themes kept in character across long scenes, and agent loops that run through to the last step.

  • GLM-5.3 Unrestricted

    glm-5.3-unrestricted · 1M · Streaming, Tool calling, JSON mode, Reasoning

    Unrestricted model built on GLM-5.3: text only, long context, tool calling and JSON mode, for long documents and hard tasks.

    input $6.00, cached input $0.60, output $10.00 per 1M tokens

  • Qwen3.5 Unrestricted

    qwen3.5-unrestricted · 256K · Streaming, Tool calling, JSON mode, Image input, Reasoning

    Unrestricted Qwen3.5 model with text and image input, tool calling and JSON mode, for everyday red-team, trust and safety and data work.

    input $2.00, cached input $0.20, output $6.00 per 1M tokens

Unrestricted does not mean without rules. Every request is covered by the Acceptable Use Policy. Acceptable Use Policy

Browse all models

/compat

Keep the client. Change the address.

Modellane answers in the request shape your code already sends. Swap the base URL and the key; messages, tools and streaming stay exactly as they are.

  • OpenAI Chat Completions: https://usemodellane.com/v1. Bearer key. Works with any OpenAI-compatible client.
  • OpenAI Responses: POST /v1/responses. Same base URL and key. The format Codex speaks.
  • Anthropic Messages: https://usemodellane.com. Key in the x-api-key header. The format Claude Code speaks.
 from openai import OpenAI
 
 client = OpenAI(
-    base_url="https://api.openai.com/v1",
-    api_key=os.environ["OPENAI_API_KEY"],
+    base_url="https://usemodellane.com/v1",
+    api_key=os.environ["MODELLANE_API_KEY"],
 )
 
 reply = client.chat.completions.create(
+    model="qwen3.5-unrestricted",
     messages=[{"role": "user", "content": "Hello"}],
 )

Read the quickstart

/privacy

We keep the bill, not the conversation.

Prompts and completions are processed in memory and discarded once the response ends. What remains is the metadata billing needs, and most of it has an expiry date.

What we never store
DataKeptHow long
Prompts and messages (Text, system prompts and tool results you send)NoDiscarded when the response ends
Images (Inline images sent to models with image input)NoDiscarded when the response ends
Completions (Everything the model writes back)NoDiscarded when the response ends

Read the privacy policy

/pricing

Pay per token, from a balance you control.

Each model lists an input, cached-input and output rate. Top up once or take a monthly plan; every request is charged from your balance when it finishes.

Prices in USD per 1M tokens
ModelAPI idContextInputCached inputOutput
GLM-5.3 Unrestrictedglm-5.3-unrestricted1M$6.00$0.60$10.00
Qwen3.5 Unrestrictedqwen3.5-unrestricted256K$2.00$0.20$6.00

Reasoning tokens are billed as output. Cached input applies when a prompt prefix is served from cache.

  • Top-ups

    Pay as you go. Add balance with a one-time payment and spend it at per-token rates. No subscription, no minimum monthly spend.

    Packages

    • $30
    • $50
    • $100
    • $200

  • Plans

    Monthly. An optional monthly plan adds a plan balance each period that is spent before your top-up balance.

    • Modellane Starter $25 / month, $30 plan balance each period · 16% off
    • Modellane Pro $50 / month, $60 plan balance each period · 16% off

Top-up credit equals the amount paid and does not expire. Plan balance is spent first and resets each period. Payments are processed by our merchant of record.

Open billing · See pricing

Full pricing

/integrations

Bring the agent you already run.

Coding agents, chat apps and SDKs connect with a base URL and a key. Every tool below has a setup guide, and the line under its name shows the endpoint it uses.

chat
/v1/chat/completions
responses
/v1/responses
messages
/v1/messages

All integrations

/faq

Questions, answered.

Who is Modellane for?
Teams whose legitimate work gets refused elsewhere: red teams and security researchers, trust and safety teams, synthetic data and ML research teams, and builders of fiction, roleplay and agent products. Security work must target systems you are authorised to test.
Which API formats are supported?
Three. OpenAI Chat Completions at https://usemodellane.com/v1, OpenAI Responses at POST /v1/responses on the same base URL, and Anthropic Messages at https://usemodellane.com with your key in the x-api-key header. Streaming, tools and JSON mode use the fields each format already defines.
Does it work with Codex or Claude Code?
Yes. Codex connects through the Responses endpoint and Claude Code through the Messages endpoint. Set the base URL and your Modellane key as shown in each integration guide, and put a Modellane model id in the model field.
How is usage billed?
Per token, from your prepaid balance. Input, cached input and output tokens each have their own rate per 1M tokens, listed in the pricing table. Each request is charged once, when it finishes.
What are cached input tokens?
When a request starts with a prompt prefix that was processed recently, the matching input tokens can be served from cache and are billed at the lower cached-input rate. Keep long system prompts and shared context at the start of the conversation to make hits more likely.
Are reasoning tokens billed?
Yes. Tokens a model spends reasoning before its answer are billed as output tokens, at the output rate.
What do you store about my requests?
Never the content. For each request we keep a record without the prompt or completion (model, key, timing, token counts, charge and status) for 180 days, then delete it.
Are there rate limits?
Yes. Each project has a limit on concurrent requests and on requests per minute. Above it the API returns HTTP 429 and you can retry after a short pause. The current limits are in the Rate limits page of the docs.
Can I get a refund?
Refunds follow our Refund Policy, linked in the footer. Write to support with your project name and the time of the payment.
What does unrestricted mean?
Unrestricted models refuse far less often than typical assistants, and we add no refusal or moderation layer of our own to any model; you apply your own product policy. It does not mean everything is allowed. The Acceptable Use Policy prohibits, among other things, sexual content involving minors, sexual content about real people without consent, help with weapons capable of mass harm, malware aimed at systems you have no permission to test, and illegal surveillance. Accounts that break it are suspended.

Your first request is three lines away.

Create a key, add balance and send your first request. The prompt is answered, then forgotten.

Get API key · Read the quickstart