Vynaris

Vynaris sits in front of the major model APIs behind one OpenAI- and Anthropic-compatible endpoint (https://api.vynaris.com/v1). You change base_url and set model to "auto"; each request starts in what Vynaris calls a quality-safe frontier pool and is routed down to a cheaper model only after the vendor says eval evidence clears a quality bar for that task. Every response carries response headers (x-vynaris-served-model, x-vynaris-cost-usd, x-vynaris-baseline-usd, x-vynaris-saved) showing exactly which model served the request and what the same tokens would have cost at the originally-requested model's list price. That per-response, exportable attribution — rather than a single blended "we saved you X%" marketing number — is the most concrete thing on the site (https://vynaris.com/docs#hosted-models).

Pay-as-you-go: you're billed provider list price for whichever model actually served the request, plus a Vynaris fee of 3% on the first $500 of usage per calendar month, then 1% beyond that; a $500 top-up activates the 1% rate immediately (https://vynaris.com/pricing, verified 2026-09-15). Prepaid top-ups start at $20 and don't expire. Fixed monthly plans convert a flat payment into rolling API credit: $20/month, $45/month, and $150/month, all with the same feature list (full model pool, receipts, no overage charges, unused credit rolls over); the $45 and $150 tiers additionally unlock the privately hosted "uncensored" models. An enterprise tier ($5,000/year) installs the gateway inside the buyer's own infrastructure, routed through the buyer's own provider accounts and rates — Vynaris never touches that compute spend and is paid only the flat annual fee for setup, patches, and updates (https://vynaris.com/enterprise).

The homepage claims 96-98% measured savings on live routed requests versus the requested model at list price, sourced to nothing but Vynaris's own traffic (vendor claim, unverified by us). To its credit, the site does not publish house benchmarks for model quality and says so explicitly — it cites three independent academic sources for the general idea of quality-preserving model routing (FrugalGPT, arxiv.org/abs/2305.05176; RouteLLM, arxiv.org/abs/2406.18665; Hybrid LLM, arxiv.org/abs/2404.14618) and states plainly that its own contribution is the per-request receipt, not the routing technique. The /models page's "worked example" (500K requests/month at Claude Sonnet 4.x pricing routed instead to DeepSeek V3) is disclosed as arithmetic on published list prices, not a benchmark or a promise about any particular buyer's traffic (https://vynaris.com/models).

Separately from routing, Vynaris directly hosts three reduced-refusal open-weight models (Qwen3.6-35B-A3B, Qwen3.8-27B, DeepSeek-V4-Flash-uncensored) at published per-token rates ($1-2 input / $5-11 output per million tokens, 128K context) and positions them for "lawful, authorized red teaming, cyber defense, and model evaluation," explicitly prohibiting CSAM, trafficking, and non-consensual sexual content (https://vynaris.com/uncensored-models). Vynaris states that prompt and output bodies for these specific model IDs are never persisted, even if transcript capture is enabled elsewhere on the account — a narrower and more specific retention promise than the general privacy policy gives routed traffic. The site is also honest that reduced refusal does not imply accuracy or safety, and that most of its "uncensored model directory" entries are research guides to community Hugging Face releases rather than models Vynaris actually hosts — only 3 of 11 listed families carry a real Vynaris model ID.

For ordinary routed traffic, prompts and completions pass through to whichever third-party provider serves the request, and each API key starts on "metadata-only" logging (Vynaris retains model, tokens, cost, latency — not content) unless the customer explicitly enables transcript capture on that key, in which case full request/response payloads are stored for dashboard debugging, quality-refund review, and — for opted-in keys — sent to a third-party routing-evaluation service (Not Diamond) purely to compare model-selection recommendations. Vynaris states it does not train on customer content itself, but a provider may do so if the customer enables provider data collection for a key. The company's own servers are US-based; routed prompts also cross into whichever region the serving provider operates in (https://vynaris.com/privacy, effective 21 August 2026).

Vynaris is a product of Wavicle.tech, a one-person, fixed-price AI consulting studio ("Built by someone who ships" — https://www.wavicle.tech/about) rather than a funded startup. The terms of service state Vynaris is "operated by a sole proprietor based in India," with disputes governed by the laws of India and the courts of Gujarat (https://vynaris.com/terms, effective 21 August 2026). The Wayback Machine's earliest capture of vynaris.com is 3 November 2025 (https://web.archive.org/web/20251103150307/http://vynaris.com/), putting the product at roughly ten months old as of this review. The site itself carries an "Early beta" badge in its header. We found no independent reviews, no third-party coverage, and no evidence of outside funding; all social presence traces to the same individual (x.com/dhrumiliso31702, linkedin.com/in/bdhrumil).

The terms of service state outright that the self-serve plan carries no uptime SLA, that upstream provider outages or degradation are passed through without Vynaris controlling them, and that total liability for any claim is capped at whatever the customer paid Vynaris in the preceding three months, excluding indirect or consequential damages (https://vynaris.com/terms). Refunds of unused prepaid balance are available on request (minus payment-processing fees); promotional credit is not refundable. There's also a "quality refund" mechanism for auto-routed requests that fail, credited as Vynaris balance rather than cash, subject to fair-use review.