Paying $1.50 per 1,000 captchas to external solver APIs eats up your margins when running 500k daily requests. By running a quantized **YOLOv8n ONNX** model for 3x3/4x4 grid classification (inference time: `14ms` on CPU!) and **faster-whisper (`tiny.en`)** for audio challenges, our self-hosted FastAPI solver hits a 94.6% pass rate at zero marginal cost. ```python from faster_whisper import WhisperModel model = WhisperModel('tiny.en', device='cpu', compute_type='int8') segments, _ = model.transcribe('captcha_audio.mp3', beam_size=1) solution = ''.join([s.text for s in segments]).strip() ```