Light
Easy tasks: formatting, short edits, quick replies.
- 1GPT-6 LunaInput $0.13Output $0.65
- 2DeepSeek V4.1 FlashInput $0.39Output $1.56
- 3Gemini 3.1 Flash-LiteInput $0.33Output $1.95
- 4Gemini 3.5 Flash-LiteInput $0.39Output $3.25
Liquid Router
Liquid Router
Liquid Router sends every request to the right model and saves your credits: low-cost models for easy work, top models for hard work.
- Model name
liquid/d1-router - OpenAI-compatible API
- 8 models in 3 tiers
- Automatic fallback
How it works
Liquid AI’s d1 model rates every request for difficulty in a single call and sends it to a model that fits.
Your requestAny prompt or agent step
d1 Routerrates difficulty
Easy tasksGPT-6 LunaSmall model · $0.10 / 1MCapability Price $$$$
Medium tasksGPT-6 SolMid-size model · $2 / 1MCapability Price $$$$
Hard tasksClaude Opus 5.5Large model · $4 / 1MCapability Price $$$$
Your requestAny prompt or agent step
d1 Routerrates difficulty
- Easy tasksGPT-6 LunaSmall model · lowest price · $0.10 / 1M
- Medium tasksGPT-6 SolMid-size model · mid price · $2 / 1M
- Hard tasksClaude Opus 5.5Large model · top capability · $4 / 1M
Stronger models cost more. Liquid Router pays for the strong one only when the task needs it.
Try an example
Pick a request and see how d1 rates it and which model answers.
Pick a request
Request
“Format this JSON config”
d1 rates the difficulty
Easy
Routed to the Light tier
- LightEasy tasks GPT-6 Luna $0.13
- StandardMedium tasks GPT-6.1 Sol $2.60
- HeavyHard tasks Claude Opus 5.5 $5.20
Price per 1M input tokens
Eight models. Three tiers.
One tier for each level of task difficulty. The first model in a tier answers, and the rest are backups, in order.
Standard
Medium tasks: summaries, reviews, multi-step work.
- 1GPT-6.1 SolInput $2.60Output $13.00
- 2Claude Sonnet 5.5Input $2.60Output $13.00
Heavy
Hard tasks: complex reasoning, tricky code, planning.
- 1Claude Opus 5.5Input $5.20Output $26.00
- 2GPT-6 AstraInput $13.00Output $65.00
USD per 1M tokens, at AI/ML API list prices.
7 tasks, routed
- 1 Generate a commit message from a diff Light GPT-6 Luna $0.0003
- 2 Convert a JSON payload to a TypeScript type Light GPT-6 Luna $0.0004
- 3 Rename a function across one file Light GPT-6 Luna $0.0021
- 4 Write a unit test for a pure function Light GPT-6 Luna $0.0007
- 5 Summarize errors in a 2,000-line log Standard GPT-6.1 Sol $0.064
- 6 Review a pull request for bugs Standard Claude Sonnet 5.5Backup: GPT-6.1 Sol was busy $0.051
- 7 Fix a race condition in a payment service Heavy Claude Opus 5.5 $0.136
Total for 7 tasks
With Liquid Router$0.25
All on Claude Opus 5.5$0.49
Saved48%
Illustrative token counts. Costs use the list prices above. Routing fee not included.
See what you’d save
Set your monthly volume and task mix. The totals update as you go.
Medium requests (Standard tier)30%
With Liquid Router$2,530
Everything on Claude Opus 5.5$9,360
You save$6,83073% less than everything on Claude Opus 5.5
Monthly estimate. Assumes 3,000 input and 600 output tokens per request, each answered by its tier’s first model. Includes one d1 routing decision per request, at 3,000 input tokens.
Why not just pick one model?
One model is either too expensive for easy work or too weak for hard work.
| Criteria | Liquid Router | One top model | One low-cost model | Your own router |
|---|---|---|---|---|
| Cost of easy tasks | From $0.13 per 1M input tokens | Full top-model price | Low | Low |
| Quality on hard tasks | Top models | Top model | Often falls short | Depends on your rules |
| Setup | Change 2 lines | Change 2 lines | Change 2 lines | Weeks of work |
| Backup if a model fails | Automatic | None | None | If you build it |
| Tasks per $1 in the example | 28 | 14 | Varies | Varies |
Swipe the table sideways to see all four options.
Switch in two lines
Keep your OpenAI-compatible code. Change the base URL and the model name. Already on AI/ML API? Change only the model name.
import osfrom openai import OpenAI client = OpenAI( base_url="https://api.aimlapi.com/v1", api_key=os.environ["AIMLAPI_KEY"],) resp = client.chat.completions.create( model="liquid/d1-router", messages=[{"role": "user", "content": "Summarize this contract"}],) print(resp.model_extra["meta"]["model"]) # the model Liquid Router pickedHighlighted code is all you change.
Questions
What do I pay for?
Two things: the d1 routing step, and the tokens used by the model that answers. Both are billed at standard AI/ML API rates and appear in your usage log.
What is d1?
A decision model by Liquid AI. It rates a request in a single call and returns calibrated probabilities with zero output tokens.
Which models can my requests go to?
Only the eight models listed on this page, in three tiers.
What happens if a model goes down?
Another provider serves the same model. If none can, the next model in the same tier answers.
Will hard tasks end up on a cheap model?
No. Requests d1 rates as hard go to the Heavy tier: Claude Opus 5.5 first, GPT-6 Astra as backup.
Can I see which model answered?
Yes. The meta.model field in every response names it. Without streaming, the x-aimlapi-model header does too.
When should I pin a model instead?
When you need the same model every time, such as for evaluations or audited workflows. You can use Liquid Router and pinned models side by side.