Stop picking models. Start routing
Model Navigator runs on GMI's own routing model, trained in-house on our benchmark. It reads each prompt and routes it to the model in your pool that answers best for the lowest cost
You pay only for the model that runs your prompt
ROUTED BY GMI'S OWN TRAINED MODEL
SCORED ON GMI'S BENCHMARK
YOUR PROMPT, RUN VERBATIM
15
Benchmarked models in the routing pool
8
Task categories
4,303
Reference prompts
Routes across 15 benchmarked models, open- and closed-source.
The most-used and strongest open- and closed-source families across real workloads, with more being benchmarked and added over time.
Anthropic
OpenAI
DeepSeek
Z.AI / GLM
Tencent / Hunyuan
MiniMax
Alibaba / Qwen
More models' benchmark coming soon in next release.
No model wins every task. So don't bet on one
Built on GMI's own benchmark across real task categories. The leader changes from one category to the next, which is why routing per prompt beats a fixed choice
A selection of the models we currently disclose; more are benchmarked internally but not yet published. Each carries a quality score and a cost score per task category
Quality scores come from GMI's internal benchmark bank; cost scores from an internal auto-eval. Figures are directional and shown for illustration, raw sources and scoring internals aren't disclosed, and cost does not imply a fixed price, which depends on your prompt
Read the prompt. Route the model
Routing runs on GMI's own machine-learning model, trained in-house on our benchmark. It reads each prompt and picks the model before any tokens are generated, at no extra cost
It starts by knowing what you're asking
Before a single token is generated, GMI's routing layer reads each prompt and compares it against our benchmark of 4,303 reference prompts with known per-model results. That comparison is what lets routing predict the best model for your specific prompt
4,303
Reference prompts in our database
8
Task types mapped
Analyze
Reads your prompt and finds the closest matches in our 4,303-prompt benchmark bank
Score
Every model has already been graded on that bank, so the router reads off each model's measured quality on prompts like yours
Rank
Weighs that quality against each model's cost, with your mode setting the balance, and ranks the models in your allowed pool
Route
The top model runs your prompt; the response returns the model that answered
THE 8 TASK CATEGORIES IN GMI'S BENCHMARK
Coding
Refactors, tests, fixes
Agent / Tool Use
Multi-step tool calls
Reasoning
Multi-step logic
Language
Multilingual & linguistic tasks
Instruction Following
Strict output formats
Knowledge
Factual Q&A
Data Analysis
Tables, trends
Math
Proofs, computation
No routing fees, no markup. Routing runs on GMI's own model, no judge model, no extra call, no added latency
Cost — leans on lighter, cost-efficient models
Balanced — quality and cost together
Quality — the strongest models by benchmark score
Allowed Model Pool — set by the workspace owner
Auto Mode — routing on/off, per workspace
One model, every prompt. That's the tax you pay
Pin one big model to everything and you overpay on the routine prompts a lighter, high-quality model handles just as well, and still loses on the ones it can't. Routing per prompt fixes both
Fixed single model
- Every prompt hits the same model
- You overpay on the easy requests
- You underperform on the hard ones
- Quality vs. spend is a manual guess
Model Navigator
ROUTED- Reads every prompt and routes it to its best-fit model
- Routes each request to its best-fit model
- Saves your strongest models for hard tasks
- Uses cost-efficient models where quality holds
Better answers. Lower cost. Measured
In GMI's internal tests, routing each prompt to its best-fit model beat a single fixed model for every request
+2.4 pts
Higher quality than only using GPT-5.5 — Quality Mode
−28%
Lower cost per task than GPT-5.5 — Quality Mode
−84%
Lower cost per task than Opus 4.8 — Balanced Mode
Based on GMI's internal benchmark across 4,303 prompts and 8 task categories, comparing 10 GMI-served models head-to-head. Routing runs within your allowed pool, drawn from the 15-model V1 pool. Quality is a benchmark score; cost is relative to fixed-model baselines. Directional; not independently audited
Quality vs cost
UP-LEFT IS BETTER ↖Always GPT-5.5
82.5% · 1.00× cost
Always Opus 4.8
81.3% · 1.38× cost
Navigator · Quality
84.9% · 0.72× cost
Navigator · Balanced
81.3% · 0.22× cost
COST PER TASK VS only using frontier model
Benchmark results are based on GMI internal evaluation and may vary by workload, prompt length, allowed model pool, and routing settings
15 benchmarked models. One routing layer
Model Navigator routes within the 15-model V1 pool, open- and closed-source alike. Define an allowed pool and route only within what your organization approves
ALLOWED MODEL POOL
Keep routing within approved models
MODEL SCOPE
Open-source, closed-source, or all models
PRICE TIER
Control the eligible model cost range
AUTO MODE
Enable or disable automatic routing
15
MODELS IN THE V1 ROUTING POOL
Open + Closed-source
MODEL FAMILIES
Admin-managed
ALLOWED MODEL POOL
Workspace-level
ROUTING SETTINGS
One endpoint. Production routing
Send requests through the dedicated Model Navigator API. It applies your workspace routing settings, picks the best-fit model, and returns the result with full routing metadata
Built for production workflows
- Dedicated routing endpoint
- Applies your workspace routing settings
- Selected model returned in routing metadata
- Authenticated with your GMI API key
Route in production. Reliably
Routing quality is only half the story. Model Navigator runs on GMI's own inference infrastructure, with the controls and failover production teams expect
Same-network inference
Models run on GMI's own infrastructure, keeping routing and generation on one low-latency network
Automatic retry & failover
If a model fails, times out, or is rate-limited, Navigator instantly reroutes to a healthy backup, so model outages rarely surface to your users
Governance built in
Admin-managed allowed model pool, model scope, and price tier, routing stays inside your policy
Full production metrics — routing latency, throughput, and error rate under load — are being benchmarked and will be published here
Stop overpaying. Start routing
Cut unnecessary model spend while keeping benchmark-backed quality across every real AI workload. Model recommendation and routing are free, for a limited time