Credits are consumed even when output has glitches or errors. Users report actual costs being 2.8x the advertised per-character rate due to failed generations and required regenerations. You pay for every attempt, not just successful outputs.

Unused monthly credits do not roll over to next month. If you have a slow month, you effectively lose the value you paid for. This catches users off guard and devalues subscriptions for inconsistent usage patterns.

Unlike Google and other providers, ElevenLabs doesn't sell token bundles. Once you run out of tokens, you must buy a bigger pricing tier entirely. No flexibility for occasional high-usage months without permanent tier commitment.

StatusGator documented 154+ outages affecting ElevenLabs users over 9 months since May 2025. IsDown recorded 74 outages averaging 8.2 per month. Multi-hour warning incidents occurred in January 2026 alone. Service reliability is a significant concern.

ElevenLabs announced deprecation of tens of voices by end of February 2026. Users who built workflows around these voices face disruption. 'With a snap of the finger those will be gone' - no migration path provided for affected voices.

Common problem: AI switches languages or accents within a single generation. A 10-minute audio starting in American English can end up British or slip into other languages entirely. Especially problematic in longer texts.

Voice cloning is poor; even after providing many samples, it sounds horrifically fake. Without professional-quality audio recordings, cloned voice sounds robotic or distorted. Quality expectations vs reality is a major gap.

Even with stability parameters configured correctly, the same voice can have subtle variations in energy, pacing, or emotional tone between calls. This inconsistency is noticeable for customer-facing applications requiring consistent brand voice.

No phone support available. Email-only responses take 5-14 days for complex issues. Users report tickets taking several days to get past automated bots. Even paying customers describe support as 'absolutely terrible.'

Hacker News users report being blocked without notice and unable to use models. Verification system 'repeatedly rejects clear, high-quality recordings without any explanation.' No clear recourse when account is locked.

When choosing a voice, you can't preview your actual text - only a tiny sample. This forces trial-and-error that consumes credits. Users must commit to a voice without knowing how their specific content will sound.

Main TTS module limited to 5,000 characters. Any changes to already-rendered content requires full new generation, not just the edited portion. A 3,000 token render must be redone entirely to correct a single AI mistake.

UI to select voices is convoluted - you must add voices to your 'voice library' first. But if library is full (10 voices on Starter plan), you can't add new voices without removing existing ones. Workflow friction for exploring voices.

App Store reviews report the app is 'still glitchy' and requires 'fighting through glitches to make it work.' App isn't connected with Apple for subscriptions - users can't find payment info in Apple settings. Dual membership + credits model confuses users.

Hacker News developers report the API is 'not very flexible.' Setting up and using API systems is not user-friendly for those without technical knowledge. Billing at 500 credits for any session under 1 minute, not based on aggregate use.