Open-source filter
Optionally limit selection to models classified as open source in the Hydra pool.
Save on every task. Tailored to your needs. With Gemini, Claude, OpenAI, DeepSeek, Qwen, Meta, Mistral, GLM, MiniMax, Kimi, NVIDIA
01 / Product
Hydra selects a suitable model for each request. Context, complexity and your model pool guide the choice.
We evaluate every model across different tasks to understand its strengths and limitations. The results help Hydra select a suitable model for your request.
Hydra weighs quality, expected response time and cost. The API returns the model selection; your application calls the selected model.
Model selection25+ models
Your requestContext · complexity
hydra
Model XModel YModel Z
Hosted in Germany
The routing engine runs in Germany.
25+ models
One routing pool for different kinds of tasks.
Always up to date.
We add evaluated new models. Your routing stays up to date automatically.
Your own model pool
Choose which models Hydra may consider for your requests. Hydra selects a suitable model from your pool.
02 / Results
High quality.
Lower costs.
Lower costs. More room to build.
≈89%
lower generation cost in our internal benchmark
Hydra D vs. GPT-5.6 Luna max · 100 planned tasks
A broad range of text tasks across different topics and complexity levels, including reasoning. Measurements cover direct responses without tool use.
Excludes coordinator, judging and operations. Historical comparison, not a savings guarantee.
Historical Hydra configurationsReference models
GLM 5.3 Flash99/100 scored
87.95 / 100; 95% interval: 82.78–92.70
DeepSeek V4 Pro 081399/100 scored
87.72 / 100; 95% interval: 82.31–92.73
Hydra D98/100 scored
85.43 / 100; 95% interval: 79.71–90.55
GPT-5.6 Luna max99/100 scored
84.43 / 100; 95% interval: 78.54–89.94
Hydra A99/100 scored
84.06 / 100; 95% interval: 77.87–90.00
MiniMax M399/100 scored
83.91 / 100; 95% interval: 78.32–89.18
Hydra H95/100 scored
83.83 / 100; 95% interval: 77.50–89.89
GPT-OSS 120B high95/100 scored
82.82 / 100; 95% interval: 76.34–88.37
GPT-5.6 Luna low99/100 scored
81.77 / 100; 95% interval: 75.55–87.73
Mean over scored tasks. Lines: 95% bootstrap intervals. Between 95 and 99 of 100 tasks were scored per configuration.
Internal measurement · no independent holdout. Hydra A, D and H are historical configurations, not results for today’s Performance and Balanced profiles.
Methodology & full data
100 planned text tasks per configuration, direct responses without tools. Scored with Fable 5.1; means include scored tasks only. The 95% intervals use 2,000 task-bootstrap samples and are descriptive. The development set was reused, and measurement waves had different conditions. This does not establish a general ranking at equivalent quality.
Costs include recorded generation attempts for 100 planned tasks. Unresolved attempt costs make the marked values lower bounds. Coordinator, judging and operations are excluded. Hydra H was measured with an open-source filter.
Source: technical white paper, 11 Sep 2026, table 1, pp. 5–6 ↗
| Configuration | Quality / 100 | 95% interval | Scored | USD / 100 |
|---|---|---|---|---|
| GLM 5.3 Flash | 87.95 | 82.78–92.70 | 99/100 | $0.11406 |
| DeepSeek V4 Pro 0813 | 87.72 | 82.31–92.73 | 99/100 | $0.39692 |
| Hydra D | 85.43 | 79.71–90.55 | 98/100 | $0.02062 |
| GPT-5.6 Luna max | 84.43 | 78.54–89.94 | 99/100 | $0.18695 |
| Hydra A | 84.06 | 77.87–90.00 | 99/100 | ≥ $0.03989 |
| MiniMax M3 | 83.91 | 78.32–89.18 | 99/100 | $0.20855 |
| Hydra H · OSS | 83.83 | 77.50–89.89 | 95/100 | ≥ $0.04837 |
| GPT-OSS 120B high | 82.82 | 76.34–88.37 | 95/100 | ≥ $0.14708 |
| GPT-5.6 Luna low | 81.77 | 75.55–87.73 | 99/100 | $0.02974 |
03 / Developers
One API.
Your application.
Send your request and a profile. Hydra returns the routing decision; your application calls the selected model.
API documentationApplication example
# Your application API settings
[hydra]
endpoint = "https://hydraroute.com/v1/answers"
api_key = "YOUR_API_KEY"
profile = "performance"
open_source = falsecurl https://hydraroute.com/v1/answers \
-H "Authorization: Bearer $HYDRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"prompt": "Summarize this report.",
"profile": "performance"
}'UTF-8TOML
Example application settings. Hydra accepts JSON API requests.
Requires approved access and your own API key.
How data is processed
The routing engine runs in Germany. The console and API proxy use Cloudflare. The console stores technical request metadata, not prompt or response content. Connected providers apply their own processing terms.
Optional pseudonymization in the native Hydra API replaces detected personal data before further processing. Responses remain pseudonymized; original values are not restored automatically.
04 / ROUTING INSIGHTS
Routing in detail
Requests by model
Model X44%
Model Y36%
Model Z20%
Understand model selection
Request shares show how usage is distributed across models.
Generation costs$0.24 USD
Requests120
per request$0.002
Make sense of costs
Put generation costs in perspective, over time and per request.
Response time · median1.8 s
Last 7 daysp50 · 1.8 s
p952.9 s
Understand response times
Consider the median alongside slower responses for a fuller picture.
Illustrative view, not live data. Generation costs and response times refer to model execution; the routing API returns the model selection.
Join the list for early access to Hydra.
Request accesshydra
Hydra / The film0:57