Optera sits in front of your model provider and cuts the cost of every request. It runs in read-only shadow mode first, measures what it would save on your own traffic, and reconciles every dollar to your real OpenAI and Anthropic invoice. You pay a slice of what it saves, or nothing.
If you build coding or agent products, your token spend isn't a rounding line anymore. It's a number the CFO knows by name, and most of it is the same waste repeated millions of times.
The whole integration is one constructor change: a new base URL plus your Optera key in one header. Your SDK, your provider key, and your code stay exactly as they are.
Swap one base_url and add your Optera key in the x-proxy-key header; your existing OpenAI or Anthropic key passes through untouched. No vendor lock-in, no rewrite.
Optera measures would-have-saved on live traffic and reconciles it to your invoice. Nothing about your requests changes yet.
Turn on the levers you trust. Lossless ones are on already; anything that could touch quality stays opt-in and eval-gated.
from openai import OpenAI client = OpenAI( api_key="sk-your-provider-key", # base_url="https://api.openai.com/v1" base_url="https://proxy.optera.dev/v1", # ← new base URL default_headers={"x-proxy-key": "YOUR_OPTERA_KEY"}, # ← your key, from the dashboard ) # Response headers now report the split: # X-Savings-Measured-USD (cache, routing, provider cache — invoice-checked) # X-Savings-Estimated-USD (compression — labeled as an estimate)
Most "optimizers" quietly rewrite your prompts and your outputs. Optera draws a hard line: the safe, measurable wins are on by default, and anything that could change a result is opt-in, quality-gated, and never touches code, JSON, or tool calls.
Identical requests return the stored response. Nothing about the prompt or the answer changes — it just doesn't get billed twice.
We place the cache breakpoints most teams get wrong, so your provider's free 90%-off prefix cache actually fires. Provable from the cached-token counts.
Send the easy calls to a cheaper model — only when it clears a quality bar on your own traffic, with automatic fallback if it ever degrades.
Trim repeated context and output waste. Off by default, never run on structured or code responses, and it won't bust the free provider cache.
Each lever takes a measured bite out of the bill. You only pay for the tokens that remain.
Optera computes cost from the provider's own usage object — the exact tokens they bill you for, with cached reads at the discounted rate. Your dashboard total reconciles to your provider bill within about a percent. If it doesn't, that's a bug, not a rounding story. Savings are split into measured (cache, routing, provider cache) and estimated (compression), and we never dress one up as the other.
The headline plan bills on the savings we prove against your invoice. No savings, no fee.
Those show you the bill or route your calls. Optera lowers the bill and proves it. We're not an observability dashboard or a gateway you configure — we're the savings, measured on your own traffic and billed on the result. The thing none of them ship is honest, invoice-reconciled, paid-on-savings.
It helps one case, and most teams place the cache breakpoints wrong or not at all. Optera makes your provider caching actually fire, then adds the levers it can't reach: a cross-request semantic cache, eval-gated model routing, and compression that's careful not to bust the free prefix cache.
Not by default. Response-side rewriting is off, and even when you turn it on it never runs on JSON, tool calls, or code. The default promise is simple: we don't change your model's answers.
No. Any optimization step that errors or runs over its time budget forwards your request unchanged. The proxy fails open by design, with a published latency budget — p99 added latency under ~15ms on a miss, or it gets out of the way.
Your provider keys pass through in memory and are never stored or logged. Prompts are stored only as much as caching and analytics require, with a metadata-only mode that keeps even hashes out of the database — and self-hosting keeps everything in your own VPC.
OpenAI and Anthropic today, including automatic translation between their API shapes, so existing clients work unchanged. Gemini and Bedrock are next.
Nothing. Shadow mode is free and read-only — you see the savings figure produced on your own traffic before you pay anything or change a line of behavior.
From the provider's reported usage, priced with a maintained price book (cached reads at the discounted rate). Cache and routing savings are exact; compression is a labeled estimate. The split shows up per request in the X-Savings-Measured-USD and X-Savings-Estimated-USD headers.
One constructor change, leave it in shadow mode, and read the number off your own traffic.
Figures shown on this page are illustrative. Your real numbers come from your own traffic in shadow mode.