Vynaris

What is Vynaris?

Vynaris is an LLM gateway compatible with the OpenAI and Anthropic client libraries that routes each request to the cheapest model its evaluation evidence certifies for the task, escalating to a frontier model when needed. It markets itself with the line "same quality, a fraction of the cost," and frames its differentiator not as routing itself but as the receipt: every response reports which model served it, what it cost, and what the direct provider price would have been. The service is in early beta at api.vynaris.com, and a separate catalogue of privately hosted "uncensored" models is offered for authorized security testing, defensive engineering, and model evaluation.

How Vynaris works

Integration is a base URL swap. OpenAI-shaped clients point at https://api.vynaris.com/v1, and Anthropic client construction uses https://api.vynaris.com, with prompts, SDKs, and tool definitions unchanged. Setting the model field to auto hands selection to the router; sending a frontier model name instead causes the request to be right-sized with a receipt.

Requests begin in a quality-safe frontier pool, and a cheaper model becomes eligible only after exact-version evaluation evidence clears a quality gate. The served model, its cost, the baseline direct price, and the percentage saved return on response headers (x-vynaris-served-model, x-vynaris-cost-usd, x-vynaris-baseline-usd, x-vynaris-saved) and are exportable as a CSV ledger. Vynaris cites FrugalGPT (Stanford, 2023), RouteLLM (LMSYS, 2024), and Hybrid LLM (ICLR 2024) as prior art, and states that it does not publish house benchmarks, relying on published research and industry-standard benchmarks for quality claims.

Main features

• OpenAI-compatible chat completions endpoint at https://api.vynaris.com/v1, with model="auto" for routing.
• Per-request receipts plus an exportable CSV ledger showing served model, provider cost, Vynaris fee, and final charge.
• Privacy-restricted providers by default, with a stated no-training commitment from Vynaris.
• Prepaid balance with no overages: credit does not expire and requests stop when the balance reaches zero.
• Privately hosted reduced-refusal models with 128K context, no prompt or output retention, and scale-to-zero capacity.
• Documented integration paths for curl, the OpenAI Python and TypeScript SDKs, the Anthropic Python SDK, Claude Code, Cursor, Hermes, OpenClaw, and Pi/other clients.

Pricing

Vynaris offers prepaid router credit and fixed monthly plans that convert payment into API credit.

Routed requests bill at the provider list price of the model that actually serves the request, plus a routing fee of 3% until monthly usage reaches $500, then 1% for the rest of that calendar month; a $500 top-up activates the 1% rate immediately. Top-ups start at $20, credit does not expire, and there are no overages.

Fixed plans add the stated amount to the balance at each renewal, with unused credit rolling over: Vynaris S at $20/month, Vynaris L at $45/month, and Vynaris XL at $150/month. Private uncensored model access begins with the L plan.

Hosted uncensored models bill at published per-token rates: Qwen3.6-35B-A3B-uncensored at $1.00 input / $5.00 output per million tokens; Qwen3.8-27B-uncensored at $1.00 / $7.00; DeepSeek-V4-Flash-uncensored at $2.00 / $11.00; and GLM-5.3-uncensored at $5.00 / $20.00, all listed with 128K context. Checkout is handled by a merchant-of-record payment partner, with available methods varying by region.

Common use cases

• High-volume agent traffic where many calls do not need a frontier model: classification, extraction, summarization, SQL generation, ticket triage, and multi-step planning.
• Cost attribution and auditing, since each response's model and price data can be exported and recomputed.
• Teams wanting a single OpenAI-compatible endpoint across multiple upstream model pools.
• Authorized red teaming and security testing using the privately hosted reduced-refusal models.

Limitations

Vynaris is in early beta. Its native /v1/messages wire format for Anthropic-shaped traffic is not yet implemented; the documentation directs those requests through the OpenAI-compatible /v1/chat/completions endpoint and describes the Claude Code environment-variable setup as preparation for future compatibility rather than a guarantee that every Claude Code call is served today. Savings vary by workload, and the baseline is provider list price, uncached. The homepage receipt sample is labelled an illustrative replay, with live routing statistics "publishing soon." Hosted uncensored models carry an acceptable-use boundary: child sexual exploitation, human trafficking or exploitation, and non-consensual sexual content are prohibited, and testing is limited to systems and data the user owns or is authorized to assess. Uncensored access requires the $45/month plan or higher, and because billing is prepaid, service stops when the balance reaches zero.

Summary

Vynaris is an early-beta, OpenAI-compatible gateway that routes LLM requests down to cheaper models when evaluation evidence supports it and escalates to frontier models otherwise, returning a per-request receipt with the served model, cost, and baseline price. It is priced as provider list price plus a 3% fee (1% above $500 monthly usage), with $20/$45/$150 monthly plans that convert to credit, plus separately published per-token rates for privately hosted reduced-refusal models. Documented caveats include an unfinished Anthropic-native endpoint, workload-dependent savings, and prepaid spend limits.