WhichuncensoredAIAPIservesabliteratedmodels?

Pricing and policy details below were read from each provider’s own pages on October 26, 2026, and Candor’s on November 7, 2026. On November 9, 2026, Candor’s API returned “The Candor developer API is paused”, so its row is kept for reference only. This market reprices often, so treat the table as a dated snapshot and re-check before you commit.

What counts as an uncensored API?

An uncensored API is a hosted endpoint that serves open-weight models whose refusal behavior has been removed or reduced, usually by abliteration or fine-tuning, and that doesn’t add its own heavy moderation layer. The word is marketing, not a standard. It tells you nothing about what the provider’s terms allow.

Three things are worth separating:

• The model. Is it an abliterated or fine-tuned variant, or a stock model with a permissive system prompt?
• The host. Does the provider run its own GPUs, or relay to someone else’s?
• The terms. Even a model that never refuses sits behind a provider contract that can ban accounts.

This page covers hosted APIs only. It does not rank them for adult content or give affiliate picks.

How do Venice, Featherless, OpenRouter and Candor compare?

Venice and Candor sell their own uncensored models per token, Featherless sells access to a very large open-model catalog by subscription, and OpenRouter is a router rather than a host. Figures are USD per 1M tokens, input / output, from each provider’s pricing page (October 26, 2026; Candor November 7, 2026).

Three notes on reading it:

• Candor lists tool calls on every model and ships an MCP server, but its developer API is paused as of November 9, 2026, so you can’t build on it today.
• Venice’s “Private” label is defined on its privacy docs as “zero data retention, contract-enforced.” The privacy page also says this relies on contracts with infrastructure partners, so it is a trust claim.
• Featherless’s Developer price is listed on its plans page as “from $50 in credits/month”, next to a separate Pay-As-You-Go option. The Chat tier is $25/month with 4 concurrent units and up to 32K context; Developer and Pay-As-You-Go list 100 units and up to 256K. Check the concurrency page before sizing an agent fleet.

Which is best for coding workloads?

The most concrete community report on this question is about tool calls. A user on Venice’s feedback board hit a 400 error using venice-uncensored in Cursor while GLM 5 worked, and a reply there says the model lacks function calling, which IDE integrations need (Venice feedback). Verify tool-call support per model ID before you pay.

We had recommended Candor for coding agents because it lists tool calls on all its models, but its developer API is paused as of November 9, 2026, so we have no pick right now. With the providers that remain, test tool calling on the exact model ID before you wire it into an agent.

No provider here publishes a coding benchmark for its uncensored models, so check the rest against what you can verify yourself. For coding, three checkable properties matter more than marketing:

1. Context length. Featherless Developer goes up to 256K. Venice’s Qwen 3.6 Plus Uncensored has a higher price tier above 256K context.
2. A stable OpenAI-style endpoint. All four expose one, so coding assistants and scripts that accept a custom base URL can point at them by changing the URL, key and model ID.
3. Per-token price at your volume. Agentic coding burns far more tokens than chat. A flat Featherless plan with unmetered tokens only applies to its Chat tier, which has 32K context.

Run your own eval on a slice of your repo before you commit. Refusal Rate has not run these models yet and will label any future run with hardware and settings.

What to check about privacy and retention

Retention labels can change under you. OpenRouter emailed users that Google-authored models stop being ZDR-eligible on October 18, 2026, because Google retains flagged prompts (quoted email, HN). One user who had set ZDR-only routing said their settings were changed automatically for those models and that they found no public announcement (HN, Oct 16).

Check where your prompts physically go, what the provider says it stores, and who else can read them. Source code is the sensitive part of a coding workload.

• Who runs the GPU? Venice says its Private mode runs on Venice-controlled GPUs or zero-retention partner infrastructure. Its Anonymized mode relays to third-party model providers, and the page tells you to assume they store your content.
• Is there a router in the middle? On OpenRouter, each provider has its own retention and training policy. OpenRouter lets you restrict requests to providers with a given data policy, per request or account-wide.
• Is there an audit? We found no independent audit of these retention claims on the pages cited here, so treat them as stated policy.

Treat all of these as stated policy. If your codebase is covered by an NDA or regulation, self-hosting is the only option that removes the trust question.

What else do developers report in practice?

Beyond the report above, community threads center on why people want these models in the first place. These are single-user reports from Hacker News, not our tests.

• Refusals on legitimate local tasks drive demand. One commenter says Qwen3.8 27B, run locally, flatly refused to operate their sonarr and radarr media setup (HN, Oct 27). Another got constant refusals while working on their own hardware and a personal firmware recompilation project, and plans to look for an abliterated model (HN, Oct 27).
• Direct vendor APIs differ from routed ones. Commenters say DeepSeek’s own API does not offer ZDR while some hosts reached through OpenRouter do (HN, Oct 17). We haven’t verified that against DeepSeek’s terms, so read them yourself.

Takeaway: test tool calling on your exact model ID, and pin your data-policy filter instead of trusting a default.

Should you self-host instead?

Self-hosting removes the provider from the threat model but moves cost and maintenance onto you. If you are on a Mac or a single GPU, you can run quantized abliterated models locally. Start with abliterated models on Ollama and see the best abliterated models in 2026 for what is worth downloading. A cost break-even between local hardware and an API belongs to self-host vs API cost, which owns that question.

Do these providers allow anything?

No. Uncensored refers to the model, not the contract. Venice’s terms bar illegal use and child exploitation and require you to follow the usage policies of third-party model providers. Featherless’s terms require compliance with the policies of third-party providers that power its models. The model’s own license also still applies; see LLM licenses and acceptable use. Use these APIs for legitimate work, such as authorized security testing or analysis of code you own.

Disclosure

Refusal Rate is published by the team that builds Candor. Nothing here is affiliate-linked, and no other provider paid for placement.

FAQ