TTS API comparison 2026
Gradium TTS vs ElevenLabs Flash v2.5 Gradium vs ElevenLabs · text-to-speech for voice agents
Gradium TTS is in the Headroom speed tier (P95 under 120 ms); ElevenLabs Flash v2.5 is in the Fast enough speed tier (P95 from 120 ms to 293 ms). Below, each measure side by side with its source; a model leads on a measure only beyond the tie rule. No combined score, no overall winner.
Figures as of October 27, 2026 · Coval, Speko, Artificial Analysis, Cekura
Side by side
| Measure | Gradium TTS | ElevenLabs Flash v2.5 |
|---|---|---|
| Speed tierCoval P95, 30 days | Headroom | Fast enough |
| Median time to first audioCoval, 30-day window | 54 ms | 191 ms |
| P95 time to first audio (slow turns)Coval, 30-day window | 104 ms | 255 ms |
| Time to first audio, board chartSpeko, US East | 55 ms | 143 ms |
| Time to first audio, full sentence (median)Cekura | 76 ms | 182 ms |
| Time to first audio, text streamed word by word (median)Cekura | -64 ms | 177 ms |
| Naturalness EloSpeko | 1,592 | 1,494 |
| Listener Elo (95% interval)Artificial Analysis | 1,174 ±13 | 1,076 ±11 |
| Numbers and dates (0 to 1)Speko | 0.27 | not rated |
| Hard passages pronounced correctly (95% interval)Artificial Analysis, human reviewers | not tested | not tested |
| Reading accuracy on hard textCekura, transcriber-scored | 79.5% | 64.4% |
| List price per 1M charactersCoval, Speko, Artificial Analysis, Cekura | $36 to $57.8 | $40 to $50 |
In bold: the model ahead on that measure. “tie”: latencies within 10 ms, or Elo scores whose 95% intervals overlap.
Where Gradium TTS is ahead
- →Median time to first audio: 54 ms vs 191 ms (Coval, 30-day window)
- →P95 time to first audio (slow turns): 104 ms vs 255 ms (Coval, 30-day window)
- →Time to first audio, board chart: 55 ms vs 143 ms (Speko, US East)
- →Time to first audio, full sentence (median): 76 ms vs 182 ms (Cekura)
- →Time to first audio, text streamed word by word (median): -64 ms vs 177 ms (Cekura)
- →Naturalness Elo: 1,592 vs 1,494 (Speko)
- →Listener Elo (95% interval): 1,174 ±13 vs 1,076 ±11 (Artificial Analysis)
- →Reading accuracy on hard text: 79.5% vs 64.4% (Cekura, transcriber-scored)
- →List price per 1M characters: $36 to $57.8 vs $40 to $50 (Coval, Speko, Artificial Analysis, Cekura)
Where ElevenLabs Flash v2.5 is ahead
On the published measures, ElevenLabs Flash v2.5 is ahead of Gradium TTS on none beyond the tie rule. This says nothing about features the benchmarks do not measure.
ElevenLabs Flash v2.5: all measurements →Frequently asked questions
- What is the difference between Gradium TTS and ElevenLabs Flash v2.5?
- Gradium TTS (Gradium) is in the Headroom speed tier (P95 under 120 ms); ElevenLabs Flash v2.5 (ElevenLabs) is in the Fast enough speed tier (P95 from 120 ms to 293 ms). Gradium TTS leads on median time to first audio, P95 time to first audio (slow turns), time to first audio, board chart, time to first audio, full sentence (median), time to first audio, text streamed word by word (median), naturalness Elo, listener Elo (95% interval), reading accuracy on hard text and list price per 1M characters; ElevenLabs Flash v2.5 leads on no measure. No combined score is computed.
- Which is faster for real-time voice agents, Gradium TTS or ElevenLabs Flash v2.5?
- On Coval's 30-day window as of October 27, 2026, the median time to first audio is 54 ms for Gradium TTS and 191 ms for ElevenLabs Flash v2.5 (Gradium TTS ahead by 137 ms); the P95, the slow turn, is 104 ms and 255 ms (Gradium TTS ahead by 151 ms). Gaps under 10 ms are ties.
- Which sounds more natural, Gradium TTS or ElevenLabs Flash v2.5?
- Speko naturalness: 1,592 for Gradium TTS, 1,494 for ElevenLabs Flash v2.5. Artificial Analysis listener Elo: 1,174 ±13 and 1,076 ±11. The two sources use different listeners and can disagree.
- How are these two text-to-speech APIs compared?
- Speed first, on Coval's median and P95 time to first audio; then voice quality (Speko, Artificial Analysis), accuracy on numbers and dates (Speko) and list price. Latencies within 10 ms and Elo scores whose intervals overlap are ties. Scores from different benchmarks are never added into one number.
- Can these numbers be checked?
- Yes. Each model page lists every measurement with its source and the date of the reading, and links to the benchmark page it comes from. The rules are on the methodology page.