Routing-efficiency benchmark · v0

Ideal routing hits 99.7% quality at a tenth of the cost.

OmnisBench measures how close an LLM routing policy gets to the ideal quality-per-dollar frontier — and publishes every response so anyone can re-grade the numbers offline, with zero API calls. This is the first run.

364 tasks · HumanEval + GSM8K 4-model pool run 2026-08-19 Apache-2.0
99.7%
task success under ideal routing (oracle)
~90%
cheaper than always using the frontier model, at parity
$0.62
cost per 1,000 requests, ideal routing
100%
of the leaderboard re-derivable from results.json
The thesis

Every policy, plotted against the frontier

Up and to the left is better: more task success, less money. The two policies on the dashed Pareto frontier aren't beaten on both axes by anything else. Ideal routing (oracle) sits top-left; brute-forcing every request through the frontier model buys no extra quality for ~10× the spend.

Cost per 1,000 requests (log-free, linear USD) vs. task success. Points on the accent-coloured frontier are non-dominated; muted points are beaten on both axes. Hover any point for detail.
The scoreboard

Leaderboard

Four policies over the same 364 tasks and the same pinned prices. oracle is the theoretical ceiling — the cheapest model that actually solved each item, chosen with hindsight. It's the target a real router aims at, not a shippable router.

PolicyWhat it does Task success$ / 1kvs frontier modelFrontier

The honest read: on this suite the cheapest model alone (gpt-5-nano) already scores 94.5%. Routing's real prize here is recovering the last ~5 points of quality while staying ~10× cheaper than the frontier — not a magic headline number.

Under the hood

Where ideal routing actually sends the work

For each task, oracle picks the cheapest model that solves it. 342 of 364 tasks are solved most cheaply by the tiniest model — routing only escalates for the hard tail. That distribution is the opportunity a live router chases.

Task success by dataset

HumanEval (164 code tasks, unit-tested) · GSM8K (200 grade-school math, exact-match).

PolicyHumanEvalGSM8K
Why you can trust it

Don't take our word — re-grade it

Most routing savings claims are unverifiable marketing. OmnisBench publishes the exact model response for every task inside results.json. omnisbench verify re-runs the graders against those responses and re-derives this entire leaderboard — quality and cost — with no API calls and no keys. Change one stored answer and it fails.

# reproduce the whole run, or just audit the published one
$ pip install -e .
$ python scripts/prepare_datasets.py       # HumanEval + GSM8K → data/
$ python -m omnisbench.cli run    --config configs/v0.yaml --run runs/mine
$ python -m omnisbench.cli report --run runs/mine
$ python -m omnisbench.cli verify --run runs/2026-08-19  # zero-API re-grade
VERIFY OK
Questions

Frequently asked

Straight answers to the questions people actually search for about LLM routing and OmnisBench.

What is an LLM router?
An LLM router inspects each request and sends it to the cheapest model that can still handle it, instead of routing everything to one expensive frontier model. OmnisBench measures how close a given routing policy gets to that ideal — the most task success per dollar.
How much can LLM routing actually save?
In OmnisBench's first reproducible run, ideal routing reached 99.7% task success at roughly 90% lower cost than always using the frontier model (Claude Opus 5). Notably the cheapest model alone already scored 94.5%, so routing's real value is recovering the last few points of quality cheaply — and every figure is re-gradable from the published results.
Is OmnisBench an open-source alternative to Weave Router or RouterBench?
OmnisBench is an open, Apache-2.0 benchmark for LLM routing efficiency — not a router product. Where vendors publish savings claims you cannot check, OmnisBench publishes the model response for every task and lets anyone re-grade the numbers offline with a single command. It complements the academic RouterBench by being live, cost-current, and continuously verifiable.
How do I verify the results myself?
Download results.json and run omnisbench verify — it re-runs the graders against the published responses and re-derives the full leaderboard (quality and cost) with zero API calls and no keys. Change one stored answer and it fails.
Which models and datasets does it cover?
v0 covers HumanEval (164 code tasks, unit-tested) and GSM8K (200 grade-school math problems) across Claude Opus 5, GPT-5, Claude Haiku 4.5, and GPT-5-nano. Additional datasets (MBPP, LiveCodeBench, MMLU-Pro) and models are config-only additions.
Read the numbers honestly

Caveats, stated up front

oracle is a ceiling, not a product

It's chosen with hindsight (cheapest model that did solve each task). No live router can be this good; the gap between a real router and this line is the real scorecard.

The cheap model is already strong

On this task mix gpt-5-nano alone hits 94.5%. The suite will grow to harder, more agentic tasks where the quality gap — and routing's value — widens.

Two datasets, four models

v0 covers HumanEval + GSM8K across opus-5 / gpt-5 / haiku-4.5 / gpt-5-nano. MBPP, LiveCodeBench, MMLU-Pro and more models are config-only additions.

Pinned, dated prices

Costs come from a committed pricing snapshot (2026-08-18), not live lookups — so the numbers are reproducible even as list prices move.