Avartha Says Its Open-Model Voice AI Tops the Leaderboard at $2.50 an Hour
Startup Avartha claims its Ava Zero voice agent, built only from open models, beats OpenAI, Google and xAI on task completion. The figures are company-reported.
A startup called Avartha introduced Ava Zero, a voice agent it says is built entirely on open models. A voice agent is an AI that can hold a spoken conversation and complete tasks, such as handling a customer call. Avartha reports it ranks first on Sierra’s τ³-Voice benchmark across six languages, with 86.0% task completion, ahead of phonic-1 (83.2%), gpt-live-1 (79.9%), grok-voice-think-fast-2.0 (71.4%) and Gemini (68.5%).
The stack is a mix of open parts: Qwen3-ASR turns speech into text, Fish Audio Drama 3 produces the voice, GLM-5.3 acts as the main model and DeepSeek 4.1 Flash handles backend work. Avartha prices it at $2.50 per hour of conversation and says that is one-half to one-third of comparable closed alternatives. It also claims its inference engine, the software that runs the model, handles up to an order of magnitude more simultaneous conversations per GPU than vLLM and SGLang.
Read the table closely. Ava scores only 25.0% on interaction, below three rivals, which suggests it is weaker at the natural back-and-forth of a conversation. On generation quality it scores 62.2%, narrowly ahead of gpt-live-1 at 61.6%. The post does not explain how the three scores combine, and it publishes no latency numbers of its own, even though speed is crucial in voice. Its concurrency claim comes with no method or hardware details, and no licence for Ava is stated.
What’s behind it: Open models (ones whose weights anyone can download) have been catching up in text, and this is an attempt to do the same in voice. Combining several specialised open models can be cheaper than a single closed one, and it lets a business keep more control. The catch is that all figures are company-reported, and a benchmark is a test, not a phone call with a frustrated customer.
What this means for you: For most people, nothing yet. If you run a small business and are thinking about an AI phone assistant, this suggests prices may fall, so it is worth comparing a few providers. Before you trust one, listen to real calls, check delays, ask which languages are truly supported and read the licence terms.
Sources
Genie 4: DeepMind's Interactive World Model Now Makes Sound Too
Google DeepMind's Genie 4, announced last week, generates playable worlds with audio and dialogue in real time. It is a research preview with a waitlist.
