Claude vs Gemini vs ChatGPT in War Games

War Games doubles as a head-to-head benchmark. Each faction and the narrator can use a different provider, so you can watch different models make strategic decisions under the same rules.

Supported providers

• Google Gemini, the default and the cheapest.
• OpenAI / ChatGPT, including any OpenAI-compatible endpoint through OPENAI_BASE_URL, so local models via Ollama, LM Studio or vLLM work too.
• Anthropic Claude.

Mix and match

AI_PROVIDER=gemini                # default for any faction without an override
NATO_AI_PROVIDER=anthropic        # NATO uses Claude
RUSSIA_AI_PROVIDER=gemini         # Russia uses Gemini
CHINA_AI_PROVIDER=openai          # China uses ChatGPT
NARRATOR_AI_PROVIDER=anthropic    # the narrator uses Claude

You only need an API key for the providers you use.

Safe by default

On startup the server asks each configured provider whether its model id exists, using the provider's metadata endpoint, which spends no tokens. A bad model id or a rejected key stops the server and names the variable to fix. If the provider is unreachable it logs a warning and starts anyway.

Keeping it cheap

A full game is at most 80 provider calls, and it defaults to small, fast models. See the cost controls.