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.
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.
$50 increments · Launches at $10,000 · August 30 deadline
Integration
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"],
)Benchmark
The numbers, not the adjectives.
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.
Prior art — routing is proven, not novel
- 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.
How it works
Route down by default. Escalate when it matters.
Point your agent at one endpoint
Swap base_url in your OpenAI or Anthropic client. Nothing else in your stack changes.
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.
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%
Receipts, not claims
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 →Who’s behind this
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 →