The Optimizer · the AI gateway that cuts the bill

Cut your agent bill.
Prove every dollar.

Point your base_url at us. We cut your LLM bill — quality-gated and measured to the dollar — and you pay only a share of what we measurably save. One line of config, your own provider keys.

free tier · one line of config · your own provider key

Already have an account? Log in →

cost · per request live model
$$$$00tokens →
baseline frontier, every call
optimized routed · pruned · stopped early
−0% measured cost, same quality bar
The integration

One line. Your SDK and keys, unchanged.

The Optimizer is a drop-in, OpenAI-compatible proxy. Change the base URL, keep your provider keys (they pass straight through). The dashboard lights up with real, attributed savings within minutes.

# before
client = OpenAI(api_key=KEY)

# after — that’s the whole change
client = OpenAI(
  api_key=KEY,
  base_url="https://rfa-labs.com/v1"
)
How it cuts the bill

A commodity gateway is table stakes. The savings live above it.

Caching and routing are commoditized. Our defensible value is the layer on top — quality-bounded, call-efficient, agent-native, and a memory that makes your savings compound the longer you run through us.

01 / gateway

Cache · route · fallback

Exact and semantic caching, model routing to the cheapest tier that holds quality, retry → fallback → circuit breakers. The commodity base — table stakes, done right.

02 / quality-bounded

Cut cost only when quality holds

Every downshift is verified before it’s trusted: prose outputs are scored by an LLM judge against your rubric, tool calls are verified structurally against your declared tools. Below the bar, we transparently fall back. We never serve you a degraded answer to save a dollar.

03 / verify-and-stop

Buy quality with the fewest calls

Spend extra reasoning only where it changes the answer, and stop the moment it’s good enough — the opposite of best-of-N spam. ~74% of best-of-16’s gain at ~1.2 samples.

04 / agent-native

Optimize whole agent runs

Memoize tool calls, prune carried context, reuse plan templates across a 30–50-step run — where agent cost actually lives. Plain prompt-caching can’t touch this.

The request pipeline

Every call walks the same path — cheapest first.

  1. 1
    cache

    the cheapest call is the one we never make

  2. 2
    trace

    make the call smaller — prune, memoize, dedupe

  3. 3
    route

    task-class → cheapest model the evidence trusts

  4. 4
    verify

    judge prose, verify tool calls — only when exploring

  5. 5
    learn

    similarity memory + trained router route new prompts directly

  6. 6
    meter

    baseline vs actual = your savings, to the dollar

Talk to us

Bring us your highest-cost agent.

We’ll prove the savings on one workflow first — your support bot, your coding agent — then expand across your traffic.