[Feature]: support of OpenRouter for access to other models

[didier-durand]

Hi,

Many Open Weoghts models are now as powerful as those of Anthropic.

So, could you, please, add support for models served by OpenRouter: we could then use Qwen, Kimi, MiniMax and compare results. (Note: OpenRouter also serves Anthropic)

I asked that to other LLM-based projects and they could add it very easily: OpenRouter supports OpenAI (and corresponding Python SDK) 100%.

Thanks in advance!

Didier

[ChrisJr404]

Took a quick spike at this and wanted to propose a shape before sending a PR.

OpenRouter exposes both an Anthropic-compatible endpoint (https://openrouter.ai/api/v1) and an OpenAI-compatible one. Since the analyzer / verifier / enhancer all already construct anthropic.Anthropic(...), the minimal-diff path is to keep the SDK and just route the base URL + api key through env vars.

Concretely:

OPENANT_LLM_BASE_URL=https://openrouter.ai/api/v1
OPENANT_LLM_API_KEY=sk-or-v1-...

When set, every anthropic.Anthropic() constructor in the codebase picks them up (pass them through AnthropicClient → SDK). When unset, behaviour is unchanged.

Two follow-on bits that aren't free:

1. Model IDs. Right now --model opus / --model sonnet resolve to literal Claude IDs in core/analyzer.py, core/enhancer.py, generate_report.py, experiment.py, etc. For non-Claude models on OpenRouter you'd need to pass things like qwen/qwen-3-coder-480b or moonshotai/kimi-k2. I'd add a third resolution alias — --model <full-id> (i.e. if the value contains a /, use it verbatim) — so users can override without us having to maintain a Qwen/Kimi/MiniMax mapping. Keeps backwards-compat for opus / sonnet.

2. Cost / token tracking. report/generator.py has hardcoded per-model pricing ({"claude-opus-4-6": {input: 15, output: 75}}). For an unknown model ID we'd default to {input: 0, output: 0} and surface a warning, which keeps the cost rollups honest rather than guessing — happy to add a MODEL_PRICING_OVERRIDE env-var for power users who want to plug in OpenRouter's pricing.

If that direction sounds right, I'll send a PR. If you'd rather a bigger refactor (a real LLM-provider abstraction with both Anthropic and OpenAI-SDK paths), happy to do that instead — just want to confirm before sinking time on it.

[didier-durand]

Hi,

Suggestion: some other services (like OpenCode) support OpenRouter by adding it as a prefix with a second / like openrouter/moonshotai/kimi-k2

Their goal is to allow use of other LLM routing services based on this prefix