AI Platform · Model Router
One routing table, not fifteen hard-coded model names
Each task declares what it needs: a reasoning budget, a latency ceiling, a fallback and whether downgrading is allowed at all. Regulated work never silently drops to a cheaper model — it queues and pages a human.
Model calls today
7,28,770
Mean cache hit
34%
policy-versioned keys
Routes live
8
each with an owner and a budget
Downgrade-blocked routes
2
compliance and transcription
Routing table
Tap a route to open its budget, fallback behaviour and policy note
| Task | Primary | Fallback | Reasoning | p50 / p95 | Cache | Calls |
|---|---|---|---|---|---|---|
| Stoxo research answer | openai/gpt-6-astra | google/gemini-3.8-flash | high | 2.8s / 9.4s | 31% | 48,120 |
| Support reply draft | google/gemini-3.8-flash | openai/gpt-6-astra | low | 0.7s / 2.1s | 58% | 92,300 |
| Compliance rewrite | openai/gpt-6-astra | openai/gpt-6-astra | high | 3.6s / 11.2s | 12% | 6,410 |
| Voice turn generation | google/gemini-3.1-flash-tts-preview | google/gemini-3.8-flash | low | 0.4s / 0.9s | 64% | 1,38,400 |
| Call transcription | google/gemini-3.5-transcribe | google/gemini-3.5-transcribe | low | 1.9s / 4.4s | 0% | 21,900 |
| Creative variant generation | openai/gpt-image-2.5-sunburst | openai/gpt-6-astra | medium | 6.2s / 14.8s | 8% | 1,840 |
| Member semantic search | google/gemini-embedding-2 | google/gemini-embedding-2 | low | 0.1s / 0.3s | 77% | 4,10,200 |
| Churn reasoning | openai/gpt-6-astra | google/gemini-3.8-flash | medium | 2.2s / 6.7s | 22% | 9,600 |
Stoxo research answer
Citations required; refuses on unsourced price targets.
- Primary model
- openai/gpt-6-astra
- Fallback
- google/gemini-3.8-flash
- Reasoning effort
- high
- Latency p50 / p95
- 2.8s / 9.4s
- Input / output per 1k
- ₹0.9 / ₹3.4
- Calls today
- 48,120
- Quality score
- 94/100
Cache hit rate31%
Calls by route today
Volume is concentrated where caching matters most
48kStoxo
92kSupport
6kCompliance
138kVoice
22kCall
2kCreative
Routing policy
The rules the router enforces, not suggestions
- 01Regulated surfaces never downgrade to a cheaper model on failure — they queue and page a human.
- 02Cache keys include the policy version, so a prompt change invalidates stale answers automatically.
- 03A 429 backs off with jitter across the whole pipeline, not per call — one shared budget for the org.
- 04Only 429 and 5xx retry. A refusal, a denial or a 400 ends the run and surfaces the reason in the UI.
- 05Every call carries a run id so gateway usage reconciles line-by-line with the FinOps ledger.