Abliteration.ai

Abliteration.ai is an AI infrastructure platform built around hosted open-weight language models with reduced default refusal behavior. It provides developers and organizations with an API that can be used as a drop-in alternative to conventional LLM providers for workloads where provider-level refusals can interfere with authorized research, security testing, AI evaluation, or synthetic-data generation.

The platform's primary model is an abliterated open-weight LLM. Abliteration is a model-editing technique that identifies and removes a refusal-related direction from a model's internal representations rather than relying on prompt-based jailbreaks. The resulting model is intended to retain general reasoning, instruction-following, multilingual, tool-use, and multimodal capabilities while reducing learned refusal behavior.

Abliteration.ai exposes its model through APIs compatible with both the OpenAI and Anthropic SDKs. Existing applications can generally migrate by changing the API base URL, API key, and model identifier rather than rewriting their application architecture. The platform supports Chat Completions, Responses API, streaming, tool/function calling, structured outputs, and multimodal inputs.

The current abliterated-model supports a large context window, streaming, tool calling, image inputs, web search, and—on supported endpoints—video inputs. The platform also offers larger reasoning-oriented models, including abliterated-model-large and abliterated-model-large-v2.

A major part of Abliteration.ai's differentiation is that it does not position unrestricted inference as unrestricted governance. Its enterprise Policy Gateway provides a separate governance layer that sits in front of the model and evaluates requests and responses according to organization-defined policies.

Policies can produce decisions such as allow, rewrite, summarize, escalate, or refuse. Organizations can version their policies, test them in shadow mode, gradually deploy them through canary rollouts, and enforce them once validated. Each policy decision can include a reason code for downstream auditing.

The Policy Gateway can integrate with enterprise monitoring and security infrastructure. Supported audit destinations include Splunk, Datadog, Elastic, Azure Monitor, S3, and custom HTTPS webhooks, allowing organizations to keep governance records in their existing observability or SIEM environments.

Abliteration.ai also provides a synthetic training-data platform. Teams can describe a dataset schema and generate preference pairs, evaluation rows, classifier examples, adversarial corpora, and other labeled datasets. Jobs can generate up to one million records and export datasets to services including Hugging Face, Kaggle, Amazon S3, Google Cloud Storage, and Azure.

This training-data capability is particularly targeted at trust-and-safety and AI security teams. Example applications include generating datasets for harassment, fraud, jailbreak, prompt-injection, deepfake, manipulation, and other classification or evaluation tasks.

The platform is also designed for AI red-teaming. Security teams can use the model to generate adversarial test corpora and probe their own AI applications, including RAG systems, tool integrations, system prompts, and agent workflows. Abliteration.ai positions this capability for authorized security testing and model evaluation rather than as a replacement for application-level security controls.

Developer integration is intentionally straightforward. Abliteration.ai supports Python, Node/TypeScript, Go and other OpenAI-compatible workflows, along with integrations for LangChain, LlamaIndex, Vercel AI SDK, Cloudflare Workers, Codex, Claude Code, and agent tooling.

The service uses token-based usage billing. Users can start with a free one-credit preview and no credit card. Paid subscriptions currently begin with the Developer plan at $20/month, followed by Growth at $50/month and Scale at $200/month. Enterprise plans provide dedicated capacity, custom routing, compliance review, security onboarding, and additional governance capabilities.

The published model rate for the standard abliterated-model is currently approximately $3 per 1 million input tokens and $3 per 1 million output tokens, while the larger models are priced at approximately $5 per 1 million tokens. Cached input is charged at a discounted rate, and web search is billed separately.

Abliteration.ai emphasizes privacy through a zero-retention-by-default architecture. The company states that prompts and outputs are not retained by default and that operational telemetry such as token counts, timestamps, and error codes is retained for billing and reliability.