Qwen-Audio-3.1: a five-model voice stack, APIs up to 95% cheaper
Alibaba's cheaper speech lineup puts a real voice agent in side-project budget. Plus ChatGPT Voice goes agentic and a GPT-6 vision bug worth re-testing.

Copy markdown
One stack for speech, up to 95% cheaper
Alibaba's Qwen-Audio-3.1 upgrades its ASR, TTS and Realtime models and cuts API prices across the lineup — up to 95% on speech recognition, ~70% on text-to-speech, and ~85% on the realtime voice endpoint. If voice pricing was the thing keeping an agent on your someday list, that math just changed.
TTS-Next and ASR-Next: the parts that unlock new projects
The two new models are the interesting ones. TTS-Next fuses an LLM with diffusion to generate speech, sound effects and background audio in a single pass — narrated scenes, not just read-aloud text. ASR-Next adds speaker diarization with timestamps plus emotion and ambient-sound detection, so meeting notetakers and call-analytics tools stop needing three stitched-together services.
ChatGPT Voice turns into a hands-free work agent
OpenAI wired ChatGPT Voice into email, calendar and Slack and let it switch across GPT-6 Astra, Sol and Luna, alongside a new ChatGPT Work mode that drafts docs, decks and sheets by voice. It's app-side for now rather than an API, but it's a clear signal of where agentic voice is headed — worth watching if you're building in that space.
Heads-up: a GPT-6 image bug that skewed your evals
OpenAI fixed an image-encoding bug that quietly degraded image understanding on GPT-6 Sol and Luna across the API and Codex, hitting computer-use flows hardest. If you benchmarked vision or shipped image features in the last few days, rerun those evals — the weak results may have been the bug, not your prompts.