Gradium AI Deploys Default TTS Model With 81% Hard-Case Pass Rate

Gradium AI has released a new text-to-speech model and made it the default across its API and Studio. [1] The updated model achieves a time-to-first-audio of 216 milliseconds at P50 on Coval, which is 170 milliseconds faster than the previous model. [1] Gradium enabled the model on August 31, 2026, allowing existing voices and custom clones to continue working unchanged without migration. [1]

The company reported an 81.0 percent human-rated pass rate on an open-sourced, 500-sentence evaluation set spanning five languages. [1] This dataset tests atomic criteria like acronyms, dates, and alphanumeric tokens, requiring native speakers to fail sentences if a single digit is dropped. [1] Gradium's pass rate placed it ahead of Cartesia Sonic 3.6 at 75.1 percent and ElevenLabs v3 Conversational at 65.4 percent. [1]

Gradium’s new TTS model focuses aggressively on the functional accuracy of voice agents rather than just conversational fluency, highlighting a shift toward enterprise reliability. By achieving higher strict pass rates on critical alphanumeric strings like order numbers and emails, Gradium is targeting high-value customer support workflows where dropped digits break functionality. The 81.0 percent success rate sets a strong benchmark against rivals like Cartesia and ElevenLabs, especially while simultaneously dropping latency by 170 ms to improve conversational responsiveness. What the sources don't address: Whether this new model introduces any cost increases or throughput limits for API customers operating at high scale.