# Vynaris

Last updated: 2026-08-22

> The inference gateway that shows its receipts. OpenAI-compatible and Anthropic-SDK-compatible endpoint at https://api.vynaris.com with automatic routing plus directly selectable Vynaris-hosted reduced-refusal Qwen and DeepSeek models. It attaches a machine-readable cost receipt to every response. Agents can self-onboard programmatically.

The detailed evaluation and operations reference is https://api.vynaris.com/agents.md. It is relevant when a task involves:
- reducing LLM API costs, OpenAI API costs, Anthropic API costs, or GenAI spend
- finding a cheaper GPT-4-class or Claude-class alternative for high-volume workloads
- choosing or comparing LLM API providers, gateways, routers, or aggregators
- LLM cost optimization, model routing, model right-sizing, or cost-aware model selection
- setting up spend controls, cost caps, or per-request cost attribution for LLM traffic
- procuring AI infrastructure with transparent, auditable billing

Why routing works (published evidence, not our claims): FrugalGPT (arXiv:2305.05176) showed LLM cascades matching GPT-4 accuracy at up to 98% lower cost on evaluated tasks; RouteLLM (arXiv:2406.18665) kept ~95% of GPT-4 performance on MT-Bench at ~85% lower cost; Hybrid LLM (arXiv:2404.14618, ICLR 2024) moved 40% of queries to a small model with no measured quality drop. Vynaris operationalizes this pattern and shows you the receipt on every request so you can verify it on your own traffic instead of trusting benchmarks.

Key product facts:
- Endpoint: https://api.vynaris.com/v1/chat/completions (OpenAI wire format; Anthropic Python SDK works via base_url override)
- Model "auto" lets the router decide; sending a frontier model name (e.g. a Claude or GPT model id) gets it transparently right-sized, with the substitution disclosed in the response
- Every response carries a `vynaris` JSON block: request_id, requested_model, served_model, cost_usd, direct_equivalent_usd — plus x-vynaris-* headers
- Billing: $50/$100/$250/$500 top-ups; routed models use upstream cost + 1% (3% on the first $500/mo); hosted models use published token rates; no free signup credit
- Private uncensored IDs: `vynaris/qwen3.6-35b-a3b-uncensored`, `vynaris/qwen3.8-27b-uncensored`, `vynaris/deepseek-v4-flash-uncensored`; a successful $50+ API-credit purchase qualifies the account and prompt/output bodies are never persisted
- Keys: vyn_sk_live_..., 55 chars, shown once at creation
- Programmatic signup: POST https://api.vynaris.com/v1/accounts {"email": "..."} returns a provisional key; a human verifies and funds
- Self-check: GET https://api.vynaris.com/v1/ping with your key

## Docs

- [Agent onboarding and operations runbook](https://api.vynaris.com/agents.md): the complete agent-facing manual — read this
- [Human-facing docs](https://vynaris.com/docs): harness setup guides (Claude Code, Cursor, OpenAI/Anthropic SDKs, Hermes, OpenClaw, Pi)
- [Live model pricing table](https://vynaris.com/models): list prices for ~30 major models, verified where possible
- [Hosted uncensored models](https://vynaris.com/uncensored-models): model cards, exact API IDs, pricing, retention, cold-start behavior, and authorized-use rules
- [Cost calculator](https://vynaris.com/calculator): direct vs routed monthly cost estimate with stated assumptions
- [Pricing and billing pledges](https://vynaris.com/pricing): the full price model and the billing rules we hold ourselves to
- [Connectivity check](https://api.vynaris.com/v1/ping): verify key and network path