The token usage do not count cached_input_tokens.

[charlietsao]

For example, this is part of session log:

{
    "timestamp": "2026-06-22T14:24:51.859Z",
    "type": "event_msg",
    "payload": {
        "type": "token_count",
        "info": {
            "total_token_usage": {
                "input_tokens": 226616,
                "cache_creation_input_tokens": 0,
                "cached_input_tokens": 176640,
                "output_tokens": 1670,
                "reasoning_output_tokens": 529,
                "total_tokens": 228286
            },
            "last_token_usage": { "input_tokens": 26701, "cache_creation_input_tokens": 0, "cached_input_tokens": 2048, "output_tokens": 283, "reasoning_output_tokens": 211, "total_tokens": 26984 },
            "model_context_window": 190000,
            "auto_compact_token_limit": 76000
        },
        "rate_limits": null
    }
}

As discussed here, the total tokens are input_tokens + cached_input_tokens. cached_input_tokens are cheap, but not free.