Gradium AI has released a new text-to-speech model that outperforms ElevenLabs and other competitors on the specific phrases that voice agents struggle with most: phone numbers, email addresses, and reference codes. The model achieved an 81.0% human-rated pass rate on a 500-sentence evaluation set of difficult real-world scenarios, compared to 65.4% for ElevenLabs v3 Conversational and 75.1% for Cartesia Sonic 3.6. The new model became the default across Gradium's API and Studio on August 31, 2026, with no migration required for existing users.

Why Does Accuracy on Hard Cases Matter for Voice AI?

Voice agents fail most often on exactly the parts of a call that matter most to customers. When a caller needs to provide their order number, callback digits, or email address, a single dropped digit or mispronounced letter can derail the entire interaction. Gradium built its evaluation set specifically to test these high-stakes scenarios, creating 500 sentences that span five languages: English, German, French, Spanish, and Portuguese. The test includes seven atomic criteria covering spelling, acronyms, alphanumeric tokens, dates, regular numbers, large and floating numbers, and email addresses. Three additional composite criteria stack these challenges together into realistic agent turns, such as processing orders, IT tickets, and insurance claims.

The scoring process was rigorous and human-centered. Independent native-speaker raters evaluated each sentence, and a sentence passed only if every element was pronounced correctly and completely. One dropped digit meant failure. Audio was loudness-normalized, the order of comparisons was randomized, and raters were capped at 40 comparisons with enforced breaks to prevent fatigue.

How Fast Is the New Model?

Speed matters as much as accuracy in voice AI. Customers expect near-instant responses, and delays create awkward silences that feel unnatural. Gradium's new model achieves a median time to first audio of 216 milliseconds on the Coval benchmark, which is 170 milliseconds faster than the model it replaces. More importantly, the model shows very little variance in performance. The interquartile range, which measures the spread between the 25th and 75th percentile response times, is just 30 milliseconds across 480 test runs. This consistency matters because callers experience the tail end of performance, not the median. A model with high variance will feel unpredictable and slow to some users, even if the average is fast.

Gradium is not claiming to be the fastest model on the market. Inworld TTS 2 posts a 166-millisecond median, and Fish Audio S2.1 Pro achieves 291 milliseconds. However, Gradium's claim is more nuanced: it offers the lowest hard-case failure rate at sub-250 millisecond first audio, with very little variance. This combination of accuracy and speed is what matters for production voice agents.

How to Deploy the New Model in Your Voice Application

• No Migration Required: Existing Gradium users can continue using their current setup without any changes. The new model is now the default, and all existing voice IDs and custom clones keep working unchanged.
• SDK Integration: New teams can install the Python SDK, point at the WebSocket TTS endpoint, and reuse existing voice IDs to get started with the new model.
• Open Evaluation Dataset: Gradium has open-sourced its 500-sentence evaluation set on Hugging Face under a CC BY 4.0 license, allowing developers to test the model independently and benchmark it against competitors.
• Bug Bounty Program: Gradium is offering 1 million credits for complete hard-case failure reports submitted on its Discord, incentivizing developers to find edge cases and help improve the model.

What Does This Mean for the Voice AI Market?

The release signals a shift in how voice AI vendors are competing. Rather than focusing solely on naturalness or speed, Gradium is emphasizing reliability on the specific tasks that matter in production environments. The company's decision to open-source its evaluation set is particularly noteworthy. By publishing the 500-sentence benchmark under a Creative Commons license, Gradium is inviting independent verification and setting a new standard for transparency in voice AI benchmarking. This contrasts with vendor-run benchmarks that are often opaque and difficult to reproduce.

The performance gap between Gradium and ElevenLabs on hard cases is significant. ElevenLabs v3 Conversational achieved only 65.4% on the hard-case evaluation set, meaning it failed on roughly one in three difficult sentences. For voice agents handling customer service, order processing, or technical support, this failure rate could translate to real customer frustration and repeat calls. Gradium's 81.0% pass rate represents a meaningful improvement, though it also suggests that even the best current models still struggle with one in five hard cases.

The timing of this release also matters. As voice AI becomes more central to customer service, technical support, and accessibility applications, the bar for accuracy is rising. Companies deploying voice agents can no longer accept models that work well on average but fail on the edge cases that customers actually encounter. Gradium's focus on hard-case performance reflects this market shift toward production-grade reliability.