AssemblyAI launches voice benchmark to cut LLM workload costs. Voice Standard Bench compares more than 50 models across quality, response time and price for transcript-processing tasks ranging from 30 seconds to 60 minutes. Published Oct 13, 2026, 5:31 PM EDT AssemblyAI launched Voice Standard Bench on Oct. 13, 2026, comparing more than 50 large language models for voice-product workloads so developers can match model size and cost to each transcript task. The benchmark evaluates information extraction, summarization, text correction and dictation cleanup across transcripts lasting about 30 seconds, two minutes, 15 minutes and 60 minutes. Results use a 0-to-100 quality scale and include each model's cost per hour of audio and median response time. Voice applications often send speech-to-text transcripts to an LLM for cleanup, classification, structured extraction or summaries. AssemblyAI's approach is to select models based on the task and transcript length instead of using one model throughout a pipeline. AssemblyAI's LLM Gateway provides access to more than 50 models through one OpenAI-compatible application programming interface. It uses the same API key as the company's speech-to-text service. Listed rates as of Oct. 8, 2026, included $0.10 per 1 million input tokens and $0.50 per 1 million output tokens for Qwen3.5 4B Fast. Gemma 4 31B was listed at $0.14 for input and $0.40 for output, while GPT OSS 20B was priced at $0.07 and $0.30. Nemotron Lightning 3.5 30B A3B was listed at $0.05 for input and $0.20 for output. One example modeled 10,000 monthly support calls lasting about 15 minutes each, with 3,200 input tokens and 300 output tokens per call. Estimated monthly costs ranged from $141 for Claude Sonnet 5 to $4.70 for GPT-6 Luna. At 1 million calls per month under the same assumptions, projected costs were $14,100 for Claude Sonnet 5, $470 for GPT-6 Luna and $568 for Gemma 4 31B. "Most of it doesn't need one," Kelsey Foster said, referring to frontier models. The benchmark counts format failures, refusals, timeouts and errors in its quality scores rather than measuring only successful responses. Companies can test the open-source benchmark using their own voice transcripts, task instructions and specialized terminology. Global routing carries no markup, while routing within the United States or European Union costs 10% more. Claude Sonnet 4.5 on LLM Gateway is scheduled to retire on Nov. 30, 2026. Simon Yoon. Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to [email protected].

We're excited to share that we’ve raised $50M in Series C funding led by Accel, our partners that also led our Series A, with participation from Keith Block and Smith Point Capital, Insight Partners, Daniel Gross and Nat Friedman, and Y Combinator.