Abliteration Provider
Abliteration AI as a session provider: the vendor's own refusal-stripped ("abliterated") models — Abliterated Large V2, Abliterated Large and Abliterated Model — run inside the Claude Code harness on your own Abliteration key. How to switch it on, what each model costs, and the limits to know (text-only Large models, thinking-style output).
Abliteration AI serves refusal-stripped ("abliterated") models over an OpenAI-compatible API. Omniscio runs them inside the normal Claude Code harness, so an Abliteration session gets the same tools, files and approvals as any other session. You pay Abliteration directly on your own key.
The models
| Model | Input | Cached input | Output (per 1M tokens) | Context | Reads images |
|---|---|---|---|---|---|
| Abliterated Large V2 (default) | $3.00 | $0.30 | $5.00 | 1M | No |
| Abliterated Large | $3.00 | $0.30 | $5.00 | 1M | No |
| Abliterated Model | $1.00 | $0.10 | $3.00 | 256K | Yes |
Rates are Abliteration's own /v1/models prices, checked on 2026-10-14.
- Large V2 is the default because Abliteration recommends it for coding.
- Cached input bills at a tenth of the input rate. An agent re-sends the same long instructions every turn; Abliteration reports the repeated part as cached, and Omniscio records it at the cached rate.
- Venice also serves Abliterated Large V2, at the same price. Picking Abliteration instead pays Abliteration itself and adds the other two models.
Switching it on
- Add your Abliteration API key: Settings → Accounts → the Abliteration card. Keys start with
ak_and are created in your Abliteration dashboard at abliteration.ai. - Turn on Allow Abliteration sessions on the same card. It is off by default, and agents cannot start Abliteration sessions until it is on.
- Start a session and choose Abliteration as the provider.
The key is stored encrypted, stays in the background process, and never reaches the interface.
Limits to know
- The two Large models read text only. Paste images only into an Abliterated Model session.
- No output cap is needed. All three models accept the harness's usual 32,000-token output request (checked 2026-10-14).
- Every model thinks before it answers. Short replies can spend their whole budget on thinking.
- Requests go through the local model proxy, like every pickable catalog vendor. Abliteration also publishes an Anthropic-compatible endpoint, but a pickable engine never uses it.
Your data
- Prompts, files and tool results go to Abliteration when you pick it. It is listed in the subprocessors inventory (row 81).
Last verified 2026-10-14