VynarisEarly beta Kimi K3Get your API key

OpenAI & Anthropic-compatible gateway

Your agent doesn’t need a frontier model for every call.

Change one base_url and every request starts with frontier capability and routes down only when current eval evidence certifies a cheaper model for that task — and the receipt says so.

vynaris · routing receiptsample
extract order fields−78%
qwen3-8b-ft412ms · $0.0004
illustrative replay · live routing stats publishing soon

New · Founding reservation

Kimi K3, direct—at half the public API price.

Private, self-hosted inference at Kimi K3’s native released precision. No router, no third-party inference API, and no prompt or completion retention.

Native MXFP4 / MXFP8$850 already reservedRefundable if not funded
Input$1.50/ 1M tokens
Output$7/ 1M tokens
Reserve from $50 →

$50 increments · Launches at $10,000 · August 30 deadline

One line. That’s the whole migration.

Keep your SDK, your prompts, your tool definitions. Swap the base URL and set model to auto.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.vynaris.com/v1",  # was https://api.openai.com/v1
    api_key=os.environ["VYNARIS_API_KEY"],
)

The numbers, not the adjectives.

96–98%
measured savings on live routed requests vs. the requested model at list price — visible on every response receipt
every request
logs which model served it, what it cost, and what the direct price was — exportable, so you can check our math

We don’t publish house benchmarks — vendor-run charts deserve the skepticism they get. Model quality claims come from published research and independent, industry-standard benchmarks (see below), routing methods follow them, and the only number we defend ourselves is the receipt on your own traffic — which you can export and recompute.

Savings vary by workload; baseline is provider list price, uncached.

  • FrugalGPT (Chen, Zaharia, Zou — Stanford, 2023): LLM cascades matched GPT-4 accuracy with up to 98% cost reduction on their evaluated tasks.
  • RouteLLM (LMSYS, 2024): learned routing kept 95% of GPT-4 performance on MT-Bench while cutting cost ~85%.
  • Hybrid LLM (ICLR 2024): query routing shifted 40% of traffic to a small model with no measurable quality drop.

Our contribution is not the idea — it’s the receipts: per-request attribution you can export and audit on your own traffic.

Route down by default. Escalate when it matters.

01

Point your agent at one endpoint

Swap base_url in your OpenAI or Anthropic client. Nothing else in your stack changes.

02

We classify, route, and escalate

Each request starts in a quality-safe frontier pool. A cheaper model becomes eligible only after exact-version eval evidence clears the quality gate.

03

Every response carries a receipt

The served model and exact cost ride back on the response headers:

x-vynaris-served-model: qwen3-8b-ft
x-vynaris-cost-usd: 0.0004
x-vynaris-baseline-usd: 0.0018
x-vynaris-saved: 78%

If you can’t verify it, we shouldn’t say it.

While other gateways bill you a blended rate and hope you don’t ask which model ran, every Vynaris response carries its receipt — the model, the cost, and what you’d have paid direct. No blended rates. No opaque routing fees. No margin you can’t see.

Per-response attribution

Every response names the model that served it and what it cost, in the headers and in your ledger. No blended averages, no hiding the routing.

Public stats page coming

Aggregate routing mix, pass rates, and savings across all traffic, published continuously. The same numbers we see are the numbers you see.

Escalation is a feature

When small models aren’t good enough, we route up to the model you asked for and the receipt reads −0%. You never pay in quality for our savings number.

Original provider pricing + 1%. Monthly plans from $5, with rollover.

See pricing →

Built by heavy users, for their own bills first.

Vynaris exists because our own agent workloads were burning frontier-model tokens on tool calls a 8B model handles fine. We built the router for ourselves, watched our own receipts, and productized what worked. No logos, no testimonials, no “trusted by” wall — none of that exists yet and we won’t fake it. What exists: a working gateway, receipts you can export and recompute, and a reply-to email that reaches the people who wrote the routing layer. If the numbers don’t hold on your workload, you’ll see it in the receipts before we do.

— Dhrumil · hello@vynaris.com · @dhrumiliso31702 · LinkedIn

Running >$10k/month in agent traffic? Enterprise runs self-hosted on your own infrastructure — your data stays yours, accumulates as a fine-tuning asset, and can route through your own provider accounts. Book 30 minutes and see your number on your own traffic.

Book 30 minutes →