FAQ

Frequently asked questions about abliteration.ai — compatibility, models, pricing, and policies.

Is abliteration.ai OpenAI-compatible?

Yes. Point any OpenAI SDK at https://api.abliteration.ai/v1. See OpenAI compatibility.

Is abliteration.ai uncensored?

Yes. The hosted models are abliterated to answer a wider range of prompts without a system-prompt jailbreak. Some requests may still receive a refusal.

Does abliterated-model refuse requests?

Yes, in some cases. Abliteration reduces refusal behavior, but results vary by prompt. We publish refusal checks and capability benchmarks for our hosted models; see models. You can also test prompts and tools specific to your application.

Can I turn off thinking / reasoning?

Yes, for abliterated-model and abliterated-model-large: set "thinking": false on /v1/chat/completions or /v1/messages. Large V2 cannot disable reasoning; the same control runs at low and hides the trace. See thinking and reasoning effort.

Is abliterated-model a reasoning model?

Yes. Use thinking: false to skip reasoning.

How is abliteration.ai different from OpenAI or Anthropic?

abliteration.ai serves unrestricted models through the OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages request shapes. Published policies can monitor or block behavior you define.

Can I use my existing OpenAI SDK with abliteration.ai?

Yes. Set the SDK’s base URL to https://api.abliteration.ai/v1 and pass your ak_... key as the API key — no code changes beyond those two values. Examples for Python, Node / TypeScript, and cURL.

Do you also support the Anthropic Messages API?

Yes. Same base URL — https://api.abliteration.ai/v1/messages. The Anthropic SDK works with api_key= (sent as x-api-key) or auth_token= (sent as bearer). See Anthropic compatibility.

Does Claude Code work with abliteration.ai?

Yes. Set ANTHROPIC_BASE_URL=https://api.abliteration.ai and ANTHROPIC_AUTH_TOKEN=ak_..., then run claude. Full setup in the Claude Code integration.

Does Claude Cowork work with abliteration.ai?

Yes. Configure Claude Desktop’s third-party inference mode (Developer → Configure third-party inference) with inferenceProvider: gateway and inferenceGatewayBaseUrl: https://api.abliteration.ai. Full setup in Claude Cowork integration.

Can I send images?

Yes — to abliterated-model, on every API surface. The text-only large models (abliterated-model-large-v2, abliterated-model-large) reject image content with a 400. See images.

Can I send video?

No. Video inputs are not currently supported and are rejected before inference. Use image inputs or extract representative frames before sending the request.

Do you support streaming?

Yes, on every API surface. Set stream: true (OpenAI) or stream: true on the request body (Anthropic). See streaming.

Are prompt cache hits guaranteed?

No. Prompt caching is automatic, but entries can become unavailable as capacity and serving conditions change or cached data is evicted. Use a stable cache routing hint to improve reuse and the response usage fields to confirm each hit. Cached-token counts are per request, not cumulative across a conversation. See prompt caching.

Do you support tool / function calling?

Yes, on every API surface. See tool calling.

What model IDs are available?

Three: abliterated-model (general-purpose, multimodal), abliterated-model-large-v2 (text-only, frontier-scale, GLM-5.3), and abliterated-model-large (text-only, the previous GLM-5.2 large model). See models.

Do you support embeddings?

No. Use a separate embedding provider alongside abliterated-model for the LLM.

Do you support structured outputs (response_format)?

No — response_format is ignored by the backend. See the compatibility matrix for the full list.

Do you support web search?

Yes, on all three API surfaces. Three different request shapes — see web search.

Do you support web fetch?

On the Anthropic Messages API only. Not on OpenAI chat completions or Responses. See web fetch.

Can I count tokens before sending?

Yes, via POST /v1/messages/count_tokens. See count tokens.

How much does it cost?

Usage-based, with separate Input, Cached input, and Output rates. Per 1M tokens, abliterated-model is $1.00 /$0.10 / $3.00; `abliterated-model-large-v2` and `abliterated-model-large` are$3.00 / $0.30 /$5.00. Billing uses one USD balance across all three models, topped up by subscription or prepaid packs. See pricing.

How do I get an API key?

Sign in to the console and create a key. Keys start with ak_. See authentication.

Do policy decision events contain my prompts?

No. Policy decision events contain metadata and per-policy results, not request or generated text. See Policy data and access.

Do your models have safety filters?

Platform protections cover child sexual exploitation and self-harm assistance. You can create policies to monitor or block additional behavior on requests and responses. If a legitimate use case is blocked, email [e-mail].

Do I need a separate endpoint for policies?

No. Policies run on the standard /v1/chat/completions, /v1/responses, and /v1/messages endpoints. Project-scoped policies apply automatically; use policy_ids for policies selected per request. See Policies in API requests.

Where do I report a bug?

Email [e-mail] or file an issue in the console.