RouterPlex’s explainer on abliteration and the prices of its abliterated routes is a pricing page with a definition attached. Abliteration, per the post, is a weight edit that removes refusal behaviour from a model, traced to the 2024 paper by Arditi and co-authors on refusal being mediated by a single direction. “Abliteration is a change to the weights after training, and it is not a clever prompt.” RouterPlex hosts five such routes, checked 3 November 2026, at $0.12 to $2 per million input tokens and $0.60 to $4 output. The cheapest is glm-4.7-flash-abliterated at $0.12 and $0.60 with a 200K context. The reasoning pick is glm-5.3-abliterated at $2 and $4, cut from $4 and $6 on 30 October, where a cold call of 40,000 input and 2,000 output tokens costs $0.088. The post is explicit that it contains no recipe, and that the terms bar illegal content, malware, weapons development and abuse material.
The figures are RouterPlex’s own, and the post says so twice: “RouterPlex sets these prices. They are not Z.ai, DeepSeek or Google list prices,” and the input rates sit above the stock models the routes come from. The same 40K-in, 2K-out mix on stock glm-5.3 costs $0.0648 against the abliterated $0.088. The claim that “most teams use them for security research” is asserted without a number and should be read as positioning. What survives is the operational detail. The $1 onboarding trial key returns HTTP 402 with the error type paid_model_required on all five routes, a paid top-up of at least $5 is required, the balance stops at zero rather than going negative, and every key can carry a hard spend cap.
The catalog has the gateway those ids go through. RouterPlex on apis.io carries five API pages. The post’s curl example hits the OpenAI-compatible RouterPlex Chat Completions API, the Anthropic-shaped base URL is the RouterPlex Messages API, and the catalog of ids the post tells you to copy from rather than guess is the RouterPlex Models API. The catalog record describes the same containment model the post sells, per-key hard spend budgets, model allowlists and optional rate limits, and adds what the post does not mention: an RFC 9727 API catalog, an OpenAPI 3.1 description, llms.txt and an anonymous read-only MCP server for model discovery. The agentic access profile counts five operations, four of them acting.
RouterPlex scores 56.1 on the Kin Score, strong, carried by access clarity at 84.2, discoverability at 75.0 and developer ergonomics at 66.1, with operational transparency at 36.8. Agent Readiness is 32.1, agent-ready, and error semantics is lit, which is the dimension this post exercises: a typed 402 body an agent can branch on is exactly what the rubric looks for, and the post and the contract agree. Well-known catalog, consent identity and agent skills are lit too. The gap is contract governance at 4.5. Three of these five rates changed on 30 October and two routes appeared on 3 November, and the contract that documents the 402 so carefully declares almost nothing about how its routes and rates change. Dry-run mode is also unlit, which matters for a product whose safety story is a hard cap that stops the run after the money is spent, not before.