Overview
workspace: production · last 30 days
Cost saved
$1,284.40
↓ 38% vs. raw
Tokens saved
9.6M
across 142k reqs
Cache hit rate
41%
semantic + exact
Spent vs. saved — by week
Optera sits in front of your model provider and trims the waste out of every request — semantic caching, model routing, prompt compression, and leak detection — before you ever pay for it.
Requests go in raw. They come out cached, routed, and compressed — you only pay for what's left.
Point your existing SDK at the proxy and pass your provider key through. Nothing else changes.
from openai import OpenAI client = OpenAI( # base_url="https://api.openai.com/v1" base_url="https://proxy.optera.dev/v1", # <- the only change api_key="sk-your-provider-key", ) resp = client.chat.completions.create( model="gpt-4o", messages=[{"role": "user", "content": "Hello"}], ) # response headers now include: X-Tokens-Saved, X-Cost-Saved
import OpenAI from "openai"; const client = new OpenAI({ baseURL: "https://proxy.optera.dev/v1", // <- the only change apiKey: process.env.PROVIDER_KEY, }); const resp = await client.chat.completions.create({ model: "gpt-4o", messages: [{ role: "user", content: "Hello" }], });
# Same request, just a different host. curl https://proxy.optera.dev/v1/chat/completions \ -H "Authorization: Bearer $PROVIDER_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "claude-sonnet-4", "messages": [{"role":"user","content":"Hello"}] }' # Works with the Anthropic Messages API shape too — # the proxy translates between formats automatically.
Set it up once. Every request after that is optimized automatically, and the savings stream into your dashboard.
Change one base URL and pass your existing OpenAI or Anthropic key through. No vendor lock-in, no rewrite.
Each request is checked against the cache, routed to the right model, compressed, and stripped of waste before it hits the provider.
Sign in to the dashboard to see dollars and tokens saved per feature, per model, and per request — in real time.
Each layer is independent and toggleable per workspace. Turn them all on, or only the ones you trust.
Near-identical prompts are matched by embedding similarity and served from cache — no second call to the provider.
Simple prompts get routed to cheaper models automatically, while hard ones stay on your flagship. You set the rules.
Bloated system prompts and redundant context are trimmed before they're tokenized — without touching the meaning.
Cut the "Sure! Here's…" preambles and decorative markdown that inflate output token counts on every response.
Long, repetitive context windows are condensed so you're not re-sending the same history at full price every turn.
Background detectors scan your traffic hourly for wasteful patterns — runaway prompts, retries, dead context — and flag them.
Every saved dollar and token is logged and attributed to the exact feature that saved it, broken down by model and time.
Speaks both the OpenAI and Anthropic API shapes and translates between them, so existing clients work as-is.
Gemini and Bedrock support are on the way. The proxy layer stays the same — only your model strings change.
Every workspace gets a live view of what it's saving — and exactly which layer earned it.
Dashboard figures shown are illustrative placeholders — your real numbers replace them on first sign-in.
Yes. You change the base_url your existing SDK points at and pass your provider key through. Your model strings, messages, and response handling stay exactly the same.
Caching requires storing request and response bodies, and analytics logs token counts. You control retention per workspace, and self-hosted deployments keep all data inside your own infrastructure. Swap this answer with your real data-retention policy before launch.
On a cache hit, requests return without ever touching the provider — faster than direct. On a miss, the optimization layers add a few milliseconds before forwarding. Net effect on most workloads is faster, not slower.
OpenAI and Anthropic today, including automatic translation between their API shapes. Gemini and Bedrock are on the roadmap. Any model your provider exposes works through the proxy.
For each request we measure the tokens and cost you would have paid sending it raw, subtract what was actually billed after optimization, and attribute the difference to the specific layer responsible. It's surfaced per request in the X-Tokens-Saved and X-Cost-Saved response headers.
Yes — the Scale tier supports self-hosted and VPC deployments so prompt data never leaves your environment. Get in touch and we'll walk through it.
Repoint one base URL, then sign in and watch the savings land.