Cartesia vs OpenAI
Cartesia Sonic 3 / Sonic 3.5 vs OpenAI gpt-4o-mini-tts for text-to-speech. TTFA: 271 ms vs 775 ms — edge Cartesia Langs: 42 vs 50 — edge OpenAI Computed from public benchmarks with dated sources; updated 2026-08-20.
Head to head
| Axis | Cartesia Sonic 3 / Sonic 3.5 | OpenAI gpt-4o-mini-tts | Edge |
|---|---|---|---|
| Price | $39 per 1M chars* | $15 per 1M chars* | — |
| ELO | 1203 | — | — |
| TTFA | 271 ms | 775 ms | Cartesia |
| Langs | 42 | 50 | OpenAI |
| streaming | ✓ | ✓ | — |
| cloning | ✓ | — | — |
* token-/credit-priced — the headline understates real per-unit cost, so no price edge is awarded.
Strengths & caveats
CartesiaLowest independently-measured time-to-first-audio in this table (271ms P50 for Sonic 3.5 in the Coval benchmark), and Cartesia's newer Sonic 3.6 — beta, API-only, released 2026-08-18 — currently leads the entire Artificial Analysis Speech Arena at 1285 ELO. Price is credit-/plan-based and the pricing page now states allowances in minutes of audio rather than characters, publishing no per-character rate at all; the ~$39/1M figure is derived from the $49/month Startup tier, so vergelijkbaar=false. Vendor latency claims (~90ms, 'sub-90ms' for 3.6) are roughly three times better than the measured 271ms P50. Sonic 3.5 slipped from about 4th to 7th on the Arena at an unchanged 1203 ELO; Sonic 3.6's much higher score is deliberately not used for this row, because 3.6 is still beta while the docs list 3.5 as stable and priced. The 42-language count carries over from the previous snapshot and was not re-verified in this refresh.OpenAICheap, developer-friendly, with steerable tone/emotion via a free-text 'instructions' parameter and tight integration into the OpenAI stack; ~11-13 built-in voices. Token-based pricing (unchanged at $0.60/1M text input tokens + $12/1M audio output tokens); the ~$15/1M-chars figure is a community estimate, so vergelijkbaar=false — note OpenAI's older tts-1 is separately listed at a real $15/1M characters, which is easy to confuse with it. Not ranked on the Arena: gpt-4o-mini-tts does not appear among the 99 leaderboard models, where OpenAI is represented only by tts-1, tts-1-hd and GPT-Realtime-2. Independent Coval TTFA is 775ms P50, the slowest in this table apart from ElevenLabs Multilingual v2. The ~50-language claim is vendor-stated and was not re-verified in this refresh. No voice cloning.
Sources
- SpeechifyAI — API pricing (TTS tiers)2026-08-20
- Artificial Analysis — Text to Speech Leaderboard (Speech Arena provider-voice, blind-vote ELO, 99 models)2026-08-20
- Coval — TTS benchmark data (TTFA P50/P95 and WER, rolling 7-day window on a pinned dataset), mirrored with methodology notes by Openbenchmarks; pulled 2026-08-20 06:30 UTC2026-08-20
- Coval — Best TTS Providers 2026 guide (successor to the 2026-05-04 benchmark post; now carries vendor-claimed latencies only, no benchmark table)2026-08-20
- ElevenLabs — API Pricing2026-08-20
- Cartesia — Pricing2026-08-20
- Google — Gemini Developer API Pricing2026-08-20
- OpenAI — API Pricing (developers.openai.com; the former openai.com/api/pricing/ URL now returns 403)2026-08-20
- MiniMax — Product Pricing (API docs)2026-06-22
- Deepgram — Pricing (Aura-2 and Flux TTS)2026-08-20
- Deepgram — TTS model docs (Aura-2 language list)2026-08-20
- MarkTechPost — Cartesia ships Sonic-3.6 (launch date, beta/API-only status)2026-08-20