autoThompson sampling picks based on learned quality posteriors. General workloads.
Model catalog
31 curated models. 8 providers.
One base_url.
Hand-curated catalog from Anthropic, OpenAI, Google, x-AI, Groq,
Perplexity, DeepSeek, and Moonshot. OpenAI- and Anthropic-compatible
endpoints. Set model="auto" and Thompson
sampling picks the best one per request.
Catalog
Eight providers, ranked on merit.
Per-model quality, latency, and success variance tracked live. The leaderboard is public to every caller via GET /v1/intelligence/rankings.
OpenAI
8 models · 27%gpt-4.1 · gpt-4o · o3 · o3-mini · …
Anthropic
7 models · 23%claude-opus-4-6 · claude-sonnet-4-6 · claude-haiku-4-5
gemini-2.5-pro · gemini-2.5-flash · gemini-1.5-pro
Groq
3 models · 10%llama-3.3-70b-versatile · mixtral-8x7b-32768
x-AI
2 models · 7%grok-3 · grok-3-mini
Perplexity
2 models · 7%sonar-pro · sonar
DeepSeek
2 models · 7%deepseek-chat · deepseek-reasoner
Moonshot
2 models · 6%kimi-k2.5 · kimi-k2.6
Auto-routing
Set model to a strategy, not a name.
auto:fastLowest p50 latency model. Real-time UX, streaming interfaces.
auto:floorCheapest model above the quality threshold. Bulk processing, classification, triage.
auto:bestHighest quality regardless of cost. Critical reasoning, code review, legal.
Prometheus stress-test
Simple tasks routed to DeepSeek at $0.000005/req. Complex reasoning routed to Claude Opus at $0.000156/req. A 31× cost spread, chosen automatically — per request, based on learned posterior, not hard-coded rules.