If you are building audio content and choosing between ElevenLabs and Gemini for text-to-speech, the numbers we collected will save you some time. We ran three tools on the same job, kids_episode.tts, across 879 calls over 90 days, and the gap between them is wider than we expected.
The three tools are eleven_v3 (elevenlabs/eleven_v3), Gemini Flash TTS (gemini/gemini-3.1-flash-tts-preview), and eleven_multilingual_v2 (elevenlabs/eleven_multilingual_v2). Only one of them is still running.
What follows covers what each call cost us, how long it took, and how often it failed. None of this is a benchmark. These are our prompts, our volumes, our 90-day window. A different job will produce different numbers.
eleven_multilingual_v2 is the cheapest and the fastest
eleven_multilingual_v2 handled 102 calls at $0.00139 per call, with a median response time of 0.9 seconds and zero failures. It is the cheapest tool in this comparison and the fastest.
Billed input was 21 tokens per call and output was 11 tokens per call. No caching was active on any of the three tools, so every token was billed at full rate.
Its last recorded use is 2026-08-22. The record does not say why it was not continued.
eleven_v3 costs 2.38 times more and runs 2.11 times slower
eleven_v3 handled 540 calls, the largest share of the three. Each call cost $0.00331, which is $0.00192 more per call than eleven_multilingual_v2 and 2.38 times its cost. Median response time was 1.9 seconds, 1.0 second slower than the fastest and 2.11 times its speed.
Billed input was 44 tokens per call and output was 26 tokens per call. No caching was used, so every billed token was a full-rate token.
eleven_v3 had zero failures across 540 calls. It is the only tool still active as of the last date in the record.
Gemini Flash TTS failed on nearly one in six calls
Gemini Flash TTS ran 237 calls. Each call cost $0.00198, which is 1.42 times the cost of eleven_multilingual_v2 and $0.00059 more per call. Median response time was 2.8 seconds, 1.9 seconds slower than the fastest and 3.11 times its speed.
The failure rate was 16.9 percent, meaning 40 of its 237 calls came back as failures on our side. The record does not say why those calls failed.
Gemini Flash TTS was first used on 2026-08-21 and last used on 2026-08-22. The record does not state a reason for stopping.
Token counts across the three tools
All three tools received the same job. eleven_multilingual_v2 billed 21 input tokens and produced 11 output tokens per call. eleven_v3 billed 44 input tokens and produced 26 output tokens. Gemini Flash TTS billed 44 input tokens and produced 105 output tokens.
None of the three tools used any caching. Every billed token was a full-rate token.
The output token count for Gemini Flash TTS, 105 per call, is the highest of the three. The record does not explain what those output tokens represent.
The limits of this comparison
These figures come from one job, kids_episode.tts, over 90 days and 879 calls. The prompts are ours, the volumes are ours, and the failure counts reflect what came back on our side, not any vendor's published uptime.
Cost per call is what we paid on our actual volume. It is not a rate card figure and should not be treated as one.
A reader running longer scripts, different languages, or higher volumes may see different costs, speeds and failure rates. This is a record of what happened here, not a general recommendation.
Which we kept
eleven_v3 is the tool still running as of 2026-08-25, the last date in the record. It costs 2.38 times more per call than eleven_multilingual_v2 and runs 2.11 times slower, but it returned zero failures across 540 calls.
eleven_multilingual_v2 was cheapest and fastest, and also had zero failures, but its last recorded use is 2026-08-22. The record does not say why it was not continued.
Gemini Flash TTS stopped after one day. Its failure rate of 16.9 percent is the clearest number in this comparison. The record does not say whether that rate was the reason it stopped.
eleven_v3 is the tool we are running. It costs more than eleven_multilingual_v2 and is slower, but across 540 calls it returned zero failures. Gemini Flash TTS failed on 16.9 percent of calls and was dropped after one day. If you are picking between these three for a similar job, the failure rate is the number to look at first.
Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.
Top comments (0)