12 models · 6 providers · sub-ms routing

Route any query to the optimal LLM

Multi-signal analysis scores every model on task match, complexity fit, cost, speed, and context capacity. No backend needed — runs entirely in your browser.

1
Classify Task
12 task categories via pattern matching
2
Score Complexity
9 signals: length, vocab, structure, domain
3
Rank Models
5 weighted dimensions + reasoning bonus
4
Route
Select optimal model with full explanation

Query Input

0 chars · ~0 tokens Cmd+Enter to route

Try an example

Routing Decision

Enter a query to see routing recommendations

The router scores 12 models across 5 dimensions and explains every decision.

The Routing Algorithm

Every routing decision follows a deterministic, explainable pipeline. No black boxes.

01

Task Classification

The analyzer runs the query against pattern libraries for 12 task categories: code generation, debugging, creative writing, analysis, math reasoning, summarization, translation, conversation, factual Q&A, system design, data extraction, and instruction following.

Each category has strong and moderate signal patterns. A query can match multiple categories — "Write a Python function to calculate eigenvalues" hits both code generation and math reasoning. The router uses both primary and secondary task types when scoring.

taskScores = { code_generation: 0.72, math_reasoning: 0.54, ... }
02

Complexity Estimation

Complexity is scored from 0.0 (trivial) to 1.0 (extremely hard) using 9 independent signals:

Length Longer queries correlate with higher complexity
Sentence Structure Multi-sentence, compound queries score higher
Multi-part Detection Numbered lists, semicolons, multiple questions
Technical Vocabulary Domain terms: "polymorphism", "idempotent", "stochastic"
Code Blocks Presence and size of embedded code
Explicit Markers "comprehensive", "nuanced", "prove" boost; "simple", "quick" reduce
Task Bias System design inherently complex; factual Q&A inherently simple
Cross-domain Queries spanning 3+ task types get a complexity bonus
Question Count Multiple question marks indicate multi-faceted queries
03

Model Scoring

Every model is scored across 5 weighted dimensions plus a conditional reasoning bonus:

Score = wtask × TaskMatch + wcomplexity × ComplexityFit + wcost × CostScore + wspeed × SpeedScore + wcontext × ContextFit + 2.0 × ReasoningBonus
Task Match w=1.2
How well does this model perform on the identified task? Uses the model's per-task benchmark scores, weighted 80% primary task + 15% per secondary task + versatility bonus.
Complexity Fit w=1.0
Each model has a complexity sweet spot [min, max]. Queries within range score 0.85-1.0. Below range (overkill) penalizes gently. Above range (underpowered) penalizes severely.
Cost Efficiency w=0.8
Estimated total cost = (input tokens × input price) + (output tokens × output price). Normalized so cheapest model = 1.0.
Speed w=0.7
Combined score: 40% first-token latency + 60% generation throughput (tokens/sec). Both normalized across all models.
Context Fit w=0.5
Checks if input + expected output fits the context window. Penalizes when utilization exceeds 50%, severely at 90%+.
Reasoning Bonus conditional
Activates for high-complexity queries or explicit reasoning requests. Boosts models with chain-of-thought capabilities: o3-mini (+0.15), DeepSeek R1 (+0.12), Opus (+0.10).
04

Priority Modes

Priority modes shift the weight distribution to favor different objectives:

Balanced
task:1.2 · complexity:1.0 · cost:0.8 · speed:0.7 · context:0.5
Max Quality
task:1.5 · complexity:1.3 · cost:0.2 · speed:0.2 · context:0.5
Max Speed
task:0.8 · complexity:0.5 · cost:0.4 · speed:1.5 · context:0.3
Min Cost
task:0.8 · complexity:0.6 · cost:1.5 · speed:0.5 · context:0.3

Routing Benchmark

Test the routing algorithm against 18 curated queries spanning all task types and complexity levels. See which models get selected and why.

Model Registry

All 12 models with their capabilities, pricing, complexity sweet spots, and per-task strength scores.

API Keys & Configuration

Configure API keys to enable the "Route & Execute" feature. Keys are stored locally in your browser only — never sent to any server except the LLM provider you're calling.

Router Weights

Fine-tune how the router scores models. These weights multiply each signal score in the final ranking.

1.0
1.2
0.8
0.7
0.5