Skip to content

Supported models

/Custom Provider

Supported models

Rixel works with supported cloud provider endpoints. Here's what we test and recommend.

Endpoint requirements

Your provider must implement:

  • POST /v1/chat/completions — OpenAI compatible
  • Streaming via SSE (stream: true) — recommended for long scans
  • Bearer auth header Authorization: Bearer $KEY

Optional but beneficial: tool calling / function calling for future structured findings.

{ "model": "gpt-4o", "messages": [...], "stream": true, "temperature": 0.2, "max_tokens": 4096 }

Compatibility matrix

ProviderExample modelsStreamingTools
OpenAI
gpt-4ogpt-4o-minigpt-4-turboo1o1-minio3-mini

gpt-4o for fixes, gpt-4o-mini for scans

YesYes
Anthropic
claude-3-5-sonnet-20241022claude-3-5-haiku-20241022claude-3-opus-20240229

Sonnet for reasoning heavy scans

via cloud adapter or OpenRouter

YesYes
Groq
llama-3.3-70b-versatilellama-3.1-8b-instantmixtral-8x7b-32768

Great for fast scans, low cost

YesLimited
OpenRouter
openai/gpt-4oanthropic/claude-3.5-sonnetmeta-llama/llama-3.3-70b

One key for many providers, model routing

YesYes
Together
meta-llama/Meta-Llama-3.1-405Bmistralai/Mixtral-8x22B

Open models at low cost

YesLimited
Mistral
mistral-large-latestmistral-small-latestcodestral-latest

Codestral for fix PRs

YesYes
Azure OpenAI
your-deployment-name

Enterprise compliance, set deployment as model name

YesYes

Recommendations by task

Security scans

Needs long context + reasoning. Use gpt-4o or claude-3-5-sonnet. Budget: gpt-4o-mini or Groq Llama 70b.

Fix generation

Needs accurate patching. Use gpt-4o, claude-3-5-sonnet, or codestral-latest for code.

Fast triage

Groq llama-3.3-70b-versatile — ~400 tok/s, cheap for initial pass.

Cost optimized

OpenRouter lets you route: Sonnet for critical/high, Haiku/Mini for low/info. One key, auto fallback.

Tool calling / structured output

Rixel parses model output as markdown + JSON for findings. Tool calling is optional but future-proofs you — models with native function calling produce more reliable location and severity. If your provider doesn’t support tools, Rixel falls back to text parsing.