01 / The route inspector
Different work.
Different routes.
Pick a sample request. Inspect the choices behind a single answer.
Extract the invoice number, vendor and total as JSON. Invoice INV-2048 Acme Studio Total: USD 840.00
Structured-output support and the quality floor narrow the pool. A single worker is enough for this example.
- Quality floor
- medium
- Route shape
- One model · direct response
- Models
- Model A
- Theoretical example cost
- $0.0006
{
"invoice": "INV-2048",
"vendor": "Acme Studio",
"total_usd": 840.00
}Illustrative routes and sample responses. Costs are freely chosen theoretical examples, not measured costs or price estimates. No request is sent. The historical benchmark below is separate.
Hydra at a glance.
Swipe to explore
Coordinator in Germany
Routing in Germany. Selected models may run with external providers.
Open-source filter
Narrow your model pool to open models.
Privacy in focus
Your privacy requirements shape which routes are eligible.
25+ models. One gateway.
Specialists for code, language and structured tasks in one pool.
Smart routing
The right model and reasoning level for your task.
02 / Follow the request
A route with
a reason.
Follow the selected example from intent to answer. Scroll to move through each decision.
The task sets the floor.
A bounded extraction task. The answer must fit a JSON schema; no planning or external tools are needed.
Your priorities guide the choice.
Structured-output support and the quality floor narrow the pool. A single worker is enough for this example.
One selected model per request.
In this illustration, a worker extracts the three requested fields into a structured response.
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
Selection starts with the models eligible for your request. The current planner uses the eligibility supplied by the application.
Weight quality, generation cost and response time to suit your workload. The selection uses development measurements for the relevant task category.
Each illustrated request uses one selected model. Multi-model routes are not currently available.
29 models. 261 task combinations.
The current development pool combines task-specific rankings with adjustable priorities. The optional portfolio filter follows model classifications and explicit self-hosting assumptions. Model weights and licence terms need to be checked before deployment.
03 / The decision record
Nothing
lost in the
black box.
Task features, selection criteria and chosen models can be traced in the local decision record. The animation shows an illustrative record.
Illustrative record · selected sample
- 01Request receivedExtract / invoice
- 02Quality floor setmedium
- 03Route selectedOne model · direct response
- 04Response checkedSchema validated
- 05Decision recordedappend-only · hash-bound
04 / Our internal benchmark
The numbers.
With their limits.
We show three Hydra configurations and six reference models on 100 internal tasks, evaluated with Fable 5.1. These September 2026 development results show how the configurations performed on this task set.
Hydra adjustment D
Select any configuration below ↘9 / 9
| ConfigurationTasks scored / 100 | Quality · 0–100Fable 5.1 | Generation costUSD / 100 planned tasks |
|---|---|---|
| Historical comparison model | 87.95 | $0.11406 |
| Historical comparison model | 87.72 | $0.39692 |
| Hydra adjustment | 85.43 | $0.02062 |
| Historical comparison model | 84.43 | $0.18695 |
| Hydra adjustment | 84.06 | ≥ $0.03989 + ? |
| Historical comparison model | 83.91 | $0.20855 |
| Hydra adjustment | 83.83 | ≥ $0.04837 + ? |
| Historical comparison model | 82.82 | ≥ $0.14708 + ? |
| Historical comparison model | 81.77 | $0.02974 |
≥ + ? marks reported generation cost with an unknown remainder. Quality bars use 0–100; cost bars use $0–$0.40. The chart below shows the descriptive confidence intervals.
How to read these results
Quality. Mean Fable 5.1 score over the scored tasks, not all 100 planned tasks. The chart shows descriptive 95% task-bootstrap intervals (2,000 replicates). The task set was reused; there is no independent holdout evidence.
Run conditions. A and the fixed controls: 9 September; D and H: 10 September 2026. Providers, timeouts, recovery and judging differed between waves.
Costs. Reported generation costs in USD for 100 planned tasks, including recorded attempts. Excludes the Coordinator, orchestration, judging and service overhead. A lower bound is not a complete price.
Comparison limits. Nine selected configurations on a direct-response workload without tools. Unequal scored coverage, historical controls and overlapping intervals prevent claims of universal savings, superiority or a Pareto frontier. These are internal scores, not Artificial Analysis rankings.
Our technical white paper. We explain Hydra’s architecture, model selection and benchmark methodology in detail, including task coverage and the limits of the results.
Read the technical paper · PDFView our benchmark chart
Open full-size chart ↗
05 / Build with the API
Your priorities.
A clear decision.
The Hydra API returns a routing decision. Your application calls the selected model separately. The console provides API keys, a playground, routing comparisons and technical request receipts.
Read the API docs ↗curl https://YOUR_HYDRA_CONSOLE_HOST/v1/answers \
-H "Authorization: Bearer $HYDRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt":"Explain PostgreSQL and SQLite.","profile":"balanced"}'Once production access is enabled: replace the host and use a console key. Hydra decides; your application runs the model. The docs include local examples.