According to CoinMeta, ElevenLabs has released a real-time voice model, eleven, which includes two versions: mini and max. mini is designed for low-cost conversations, while max is aimed at customer service and complex tasks that require continuous tool invocation. Both models support over 90 languages and more than 17,000 different voices. In a test with 1,000 voice recognition questions, max achieved a correct rate of 98.8% and started speaking in an average of 0.57 seconds. qwen, audio, realtime, and plus also had a high correct rate of 99.2%, but it took them an average of 1.54 seconds to start speaking. The new model has begun to be rolled out through elevenagents, realtime, and api, and is gradually becoming available to more users. api is compatible with the openai websocket protocol, allowing developers to reuse existing projects that use sdk. Projects that use webrtc will need to adjust their connection methods first.