Alibaba's Qwen Audio 3.1: five voice models, prices cut up to 95%
Alibaba's Qwen team shipped a complete audio stack today — and repriced it like it wants to own the category. Also in this brief: a Claude Code privacy quirk that got confirmed and fixed within hours, and AI agents getting a leash for the mainframe.
Alibaba's Qwen team has turned its audio research into a full product line: Qwen-Audio 3.1 bundles five speech models covering recognition, synthesis and real-time conversation, with price cuts of up to 95%. The lineup spans a speech-recognition model that auto-cleans filler words and repetitions, an ASR-Next variant that separates speakers, tags timestamps and picks up emotions and background noise, two text-to-speech generators, and a real-time duplex model that keeps listening while it talks and interrupts instantly. The next-gen TTS pairs a language model with a diffusion generator to produce voice, sound effects and background audio in a single pass; the real-time model slows down and softens when it hears a low mood, per Qwen. Prices fall up to 95% on ASR, about 85% on Realtime and about 70% on TTS — the new ASR model is listed at $0.15 per million input tokens on QwenCloud, announced alongside Alibaba Cloud's Apsara conference.
The strategy reads plainly: voice is getting the token-price treatment that text got two years ago. We noted the same play when Qwen made its existing voice model dramatically faster — Same Qwen3 voice model, 14x faster — hosting is the product now. Coverage from ifeng Tech adds detail the announcement skipped: 30 languages and 16 Chinese dialects, with roughly 160 milliseconds to the first streamed token. One caveat — the models are API-only so far; no open-weights release has been announced, and commenters are already asking.
A researcher found Claude Code only reads AGENTS.md when telemetry is on — and Anthropic called it "a fully human error" and shipped the fix the same day. Developer Przemysław Szypowicz measured that the AGENTS.md loader sits behind a remote feature flag: with telemetry or "nonessential traffic" disabled, the file is silently skipped — no warning, and the model never sees your instructions. The same gap hit Bedrock, Vertex and third-party gateway users, where the flag could never resolve. An Anthropic engineer confirmed on Hacker News that it was a rollout artifact, fixed in v2.1.281. The lasting point: a remote switch could quietly disable reading a local file, so anyone running agents in CI with traffic locked down should double-check their instructions were ever loading at all.
Rocket Software is putting governed agents on the mainframe: EVA 2.0 adds PlanGuard, a policy checkpoint that reviews an agent's proposed action before it executes. PlanGuard checks the caller, request, tool and operational context against policy, then grants a temporary execution identity scoped to the approved task — revoked when the work is done — while working alongside existing mainframe security managers like RACF and writing a tamper-evident, hash-chained audit trail. Rocket says a pilot at a South American financial institution root-caused in under a day a problem that had taken three weeks manually; the company's 3.2x ROI estimate is its own analysis, not independently verified. Agent governance is becoming its own market — Cyera raises $400M to be the trust layer for enterprise AI agents.
What to watch: whether Qwen open-weights the audio models, and whether invocation-time approval like PlanGuard's becomes the enterprise template before agents touch production systems.
Voice APIs, a remote kill switch and a mainframe with a conscience — which of these touches your work first? Tell us in the comments.
Sources: GitHub issue #95690 (Anthropic)
