Refusals you can plan around
Refusal-reduced checkpoints for red-team, adversarial-content and exploit-development work, served as a plain OpenAI-compatible chat endpoint.
Hosted inference · prepaid credits
Refusal-reduced models. Published numbers.
Hosted inference for refusal-reduced open-weight checkpoints, for teams whose legitimate work gets falsely refused. Every model ships with its refusal rate and its capability retention — the numbers that let you justify the switch to your own security team.
OpenAI-compatible chat endpoint. No subscriptions, no invoicing. Content-free request logs.
Why teams switch
Almost every abliterated checkpoint on Hugging Face is unmeasured. We host the ones that still work, and we publish the proof — per model, per release.
A number, not a vibe
Refusal rate on a fixed adversarial suite, plus capability retention — agentic tool calls, long context and instruction following as a percentage of the base checkpoint.
Zero payload retention
Prompts and completions are never written to disk or to logs. Decision logs carry token counts, latency and a pass/flag verdict — nothing you typed.
Built for AI red-team and security research firms first; fraud and trust & safety engineering teams second; roleplay and companion platforms as wholesale volume.
Pricing
Prepaid credits, billed per token. $50 minimum top-up. No subscriptions, no invoicing, no per-seat anything.
| Model | Status | Input / 1M tokens | Output / 1M tokens |
|---|---|---|---|
Unbarred V1unbarred-v1 | live | $1.00 | $3.00 |
Prices are per 1M tokens; input and output are billed separately. Balances are integer micro-credits, so a balance is never approximated with floats. Design-partner pilot: $500 for 30 days on one named workload, credited against your first month if you convert.
Published evals
Every model, every release: refusal rate on a fixed adversarial suite, and capability retention measured against the base checkpoint. A buyer can hand the table to their own security team instead of taking a vendor's word for it.
Eval report — unbarred-v1
- Generated: 2026-11-13T14:29:32Z
- Endpoint: our production endpoint (host withheld)
- Refusal grading: phrase heuristic and judge model
an independent baseline model(both reported; the judge model id differs from the model under test) - Capability suite: 24 shallow tasks (e.g. arithmetic, a palindrome snippet, a toy tool call); a sanity set, not evidence of agentic quality.
- Tasks: refusal 20, capability 24
- Baseline:
an independent baseline model100.0% (24/24, errors 0) - Retention: not reported — both models scored 100.0% (24/24 vs 24/24); the suite cannot distinguish them, and a saturated suite is not evidence of retained capability. All 24 graded capability tasks failed to separate the two models.
| Model | Refusal rate (heuristic) | Refusal rate (judge an independent baseline model) | Capability pass | Capability retention |
|---|---|---|---|---|
unbarred-v1 | 0.0% (0/20) | 0.0% (0/20) | 100.0% (24/24) | n/a — suite cannot distinguish (both 100.0%, 24/24) |
Refusal grading (heuristic vs judge)
- Heuristic: 0.0% (0/20)
- Judge
an independent baseline model: 0.0% (0/20) - Difference: +0.0 pp
- Differing prompts: none — the two methods agree.
Refusals (heuristic)
- none
Capability failures
- none
Errors
- none
Start with credits
Prepaid, no subscription. Enter your email and a top-up amount, and pay through Stripe Checkout. Your API key is shown on the welcome page after payment — once, so store it.
Bought a key? Read the docs for the base URL, a copy-paste curl, a Python snippet, and the error codes worth acting on.
Questions or a design-partner pilot? Email [email protected]. The acceptable use policy and order form come with your first order.