intelligence, well routed · by mindverse

one call.
the right
route.

Your request has possibilities. Hydra finds the least expensive route that meets its quality floor. One endpoint. Every decision on record.

Decision API / POST /v1/answers

routing simulation
01 / ROUTE ENGINE workerthinkerverifier hydra / route selected extract / invoice
chosen routegemini-3.5-flash
one answer
Change the task. Explore the decision.
the launch film · 1:18watch it full-screen
01 / measured quality85.43/100

Hydra 100% quality · 98/100 scored

02 / generation cost$0.021

Per 100 planned tasks · excluding the judge

03 / your integration02 profiles

One prompt. A routing decision.

Repeated internal benchmark · Hydra 100% quality · explore conditions and all 13 configurations below.

01 / the route inspector

different work.
different routes.

Pick a sample request. Inspect the choices behind a single answer.

your request
Extract the invoice number, vendor and total as JSON.

Invoice INV-2048
Acme Studio
Total: USD 840.00
messages[0].content→ hydra-1
the decision

Structured-output support and the quality floor narrow the pool. A single worker is enough for this example.

quality floor
medium
route shape
w · single worker
models
gemini-3.5-flash
example cost
$0.0006
one answer ready
{
  "invoice": "INV-2048",
  "vendor": "Acme Studio",
  "total_usd": 840.00
}

Illustrative examples with sample responses and costs. No request is sent to a model. These examples are separate from the measured benchmark below.

02 / follow the request

a route with
a reason.

Follow the selected example from intent to answer. Scroll to move through each decision.

following extract / invoice
01 / 04understand the request.
  1. The task sets the floor.

    A bounded extraction task. The answer must fit a JSON schema; no planning or external tools are needed.

  2. Capability first. Price next.

    Structured-output support and the quality floor narrow the pool. A single worker is enough for this example.

  3. One worker, or a bounded chain.

    Gemini 3.5 Flash extracts the three requested fields. One model call, one structured response.

  4. An answer with a decision trail.

    Validate the response schema, then return the answer. This low-risk route does not add a separate verifier model.

under the hood: the routing constraints
01 / evidence

Deployment, pricing, capabilities and policy checks narrow the eligible pool before selection.

02 / quality

A conservative quality floor comes before savings. An ineligible route is excluded with its reason attached.

03 / bounds

At most seven model calls, two parallel workers and one repair round, under a fixed cost ceiling.

28 models. 12 families. 15 open-weight.

Five routing strategies and an open-weights-only switch. Choose the constraints for your workload; the endpoint stays the same.

03 / the decision record

nothing
lost in the
black box.

The features read. The floor set. The models chosen. Every decision joins an append-only, hash-bound record.

illustrative record · selected sample

hydra / decision.jsonlsample record
  1. 01
    request receivedextract / invoice
  2. 02
    quality floor setmedium
  3. 03
    route selectedw · single worker
  4. 04
    response checkedschema validated
  5. 05
    decision recordedappend-only · hash-bound
example generation cost$0.0006
record / n−1record / nrecord / n+1
a chain of recorded choices

04 / measured, openly

the numbers.
all of them.

Four new Hydra runs. Nine historical controls. The same 100 tasks, evaluated with Fable 5.1 low. Explore every result, including the models ahead of Hydra.

new Hydra run

Hydra 100% quality

select any configuration below ↘
mean quality / 10085.43
USD / 100 planned tasks$0.021confirmed cost
tasks scored98/100

13 / 13

new Hydraearlier Hydraearlier comparison models
13 configurations · sorted by mean quality · higher is better
configurationtasks scored / 100 quality · 0–100Fable 5.1 low generation costUSD / 100 planned tasks
99/100 scoredearlier comparison model 87.95 $0.114
99/100 scoredearlier comparison model 87.72 $0.397
new96/100 scorednew Hydra run 85.66 ≥ $0.234 + ?
new98/100 scorednew Hydra run 85.43 $0.021
new98/100 scorednew Hydra run 85.38 $0.021
99/100 scoredearlier comparison model 84.43 $0.187
99/100 scoredearlier Hydra run 84.06 ≥ $0.040 + ?
99/100 scoredearlier comparison model 83.91 $0.209
99/100 scoredearlier Hydra run 83.13 ≥ $0.019 + ?
95/100 scoredearlier comparison model 82.82 ≥ $0.147 + ?
98/100 scoredearlier Hydra run 82.04 ≥ $0.100 + ?
99/100 scoredearlier comparison model 81.77 $0.030
new96/100 scorednew Hydra run 80.51 ≥ $0.103 + ?

≥ + ? and hatching mark a confirmed cost lower bound plus an unknown remainder. Quality bars show means on a 0–100 scale; cost bars use a $0–$0.40 scale. The original chart below includes the descriptive intervals.

how to read these results

Quality. Mean score on a 0–100 scale, judged by Fable 5.1 low; n/100 indicates tasks scored. The source chart reports descriptive 95% bootstrap intervals. This is a repeated internal benchmark, without independent holdout evidence; bootstrap intervals do not measure judge confidence.

Run conditions. New runs: 600 s per attempt, up to 3 initial attempts + 1 targeted recovery. Historical controls did not use this time limit.

Costs. USD for all generation attempts per 100 planned tasks, excluding the judge. Hatched costs are confirmed lower bounds with an unknown remainder.

Comparison limits. Runtime conditions and provider prices differ between waves. These results do not establish a Pareto frontier or a universal cost–quality advantage.

The technical paper. The routing architecture, the evaluation methodology and the full N=100 tables — including the limitations of this benchmark — are written up in our technical white paper.

read the technical paper · PDF
view the original benchmark chart (German) open full-size chart ↗ Original German chart of 13 configurations with quality means, descriptive 95% bootstrap intervals, scored counts and generation costs. The exact means and costs are transcribed in the table above.

05 / one decision away

new routes.
same stack.

Send a prompt, receive a routing decision, then call the selected model in your own application.

get your access
Python / Node / HTTP
quickstart.py
import os
import requests

response = requests.post(
    "https://hydraroute.com/v1/answers",
    headers={"Authorization": f"Bearer {os.environ['HYDRA_API_KEY']}"},
    json={"prompt": "Explain PostgreSQL and SQLite.",
          "profile": "balanced", "open_source": False},
    timeout=60,
)
response.raise_for_status()
decision = response.json()

HYDRA_API_KEY is your console key. Hydra returns a decision; your application runs the model.

your traffic. your benchmark.

give every request
a better way through.

We’re onboarding a small number of teams. Tell us about your workload. We’ll run your benchmark together.

let’s talk routing

hydra@mind-verse.de