What does it cost to read a rabbit aloud?

Reading aloud is an unusually unforgiving place to save money. A voice that sounds convincing for a greeting can become tiring across a chapter. A misplaced pause can change the sense of a sentence. Still, narration is something a reader might use for hours, so a small difference in the price of each paragraph eventually matters. We wanted something more useful than a list of prices: the same words, spoken by the models we are considering, with the bill beside the play button.

We used the third paragraph of Chapter I of Lewis Carroll’s Alice’s Adventures in Wonderland. It gives the voices a parenthetical aside, an exclamation, a long sentence and a rabbit in a hurry. Every request contains the same passage below. These are single takes, with their settings exposed, so you can listen for the differences without taking our word for how any of them sounds.

The reference is Green Room’s current ElevenLabs Flash v2.5 configuration, using Jessa. Alongside it are SpeechifyAI Simba 3.2, Inworld Realtime TTS-2 Flash, Murf Falcon 2, and Google’s Gemini 3.8 Flash-Lite TTS and Gemini 3.8 Flash TTS. The Google pair both use Gacrux: identical input, voice and PCM output settings, with only the model changed and no extra delivery instruction. That makes the pair a closer comparison than the other voices, although it is still just one generation each.

{
  "model_id": "eleven_flash_v2_5",
  "voice_id": "yj30vwTGJxSHezdAGsv9",
  "output_format": "mp3_44100_128",
  "voice_settings": "provider defaults, exactly as current app"
}

{
  "voice_id": "imogen_32",
  "model": "simba-3.2",
  "output_format": "wav_48000"
}

{
  "voiceId": "Cordelia",
  "modelId": "inworld-tts-2-flash",
  "audioConfig": {
    "audioEncoding": "WAV",
    "sampleRateHertz": 48000,
    "speakingRate": 1
  },
  "temperature": 1,
  "enhanceGeneration": true,
  "timestampType": "WORD"
}

{
  "model": "google/gemini-3.8-flash-lite-tts",
  "voice": "Gacrux",
  "response_format": "pcm"
}

{
  "model": "google/gemini-3.8-flash-tts",
  "voice": "Gacrux",
  "response_format": "pcm"
}

{
  "voiceId": "Theo",
  "model": "falcon-2",
  "locale": "en-UK",
  "style": "Narration",
  "format": "WAV",
  "sampleRate": 24000,
  "channelType": "MONO",
  "rate": 0,
  "pitch": 0
}

Starting a sample pauses the others. All published samples use the same listening format: mono MP3 at 96 kbps and 24 kHz, converted from the providers’ original outputs. We matched whole-recording RMS volume, including pauses, to -23.10 dBFS using constant gain. The decoded MP3s are within 0.02 dB of that target, with at least 1 dB of sample-peak headroom. There is no dynamic compression or limiting; RMS matching does not guarantee identical perceived loudness. Request time is the full request and download, not the delay until a listener could first hear streamed speech.

On this paragraph, Murf Falcon 2 has the lowest recorded projection at $0.58 per audio hour, compared with $3.02 for our current ElevenLabs setup. That is 81% lower on the cost bases shown below. This is a cost observation about these takes; it does not establish a listening-quality winner.

The two cost labels matter. “Measured billing” means a returned OpenRouter generation receipt. “Estimated usage” means a published provider rate applied to the recorded character count, with the exact assumption in the card. An estimate is not an invoice: subscriptions, included allowances and negotiated rates can change what someone pays. We have not treated a missing price as zero, or counted temporary allowances as the lasting cost of running narration.

The hourly column uses each recording’s actual duration: sample cost multiplied by 3,600, divided by its length in seconds. This is useful for thinking about listening, but pace is a confounding factor. A slower reading produces more audio minutes from the same text and can therefore look cheaper per hour without being cheaper per book. The sample-price column holds the text constant; the hourly column tells us what that particular performance would cost at its particular pace.

There is another requirement that a play button cannot show. Green Room needs to know which words are being spoken so the reading highlight can follow along. Audio alone leaves that connection missing. Each card records the alignment evidence returned by its request. A good-sounding model without usable timestamps would need an additional alignment step, and its cost and reliability would belong in the comparison too. A provider advertising timestamps somewhere in its API is not enough; the narration route we actually use must supply them.

We have kept the app’s narration model unchanged while making this comparison. The next decision needs longer passages, dialogue, difficult names and checks that no words disappear or repeat. Voice preference also deserves room: an economical take that you would stop listening to after ten minutes has missed the point. For now, the recordings and their costs give us a concrete place to start choosing.

Captured 2026-09-24. All prices are USD. Settings and provider sources are linked with each recording. Source passage SHA-256: fe13915de3945aff9a34fd49a1b093e065ba6943f138fc5af530094e27865843.