AI News: September 24, 2026

1. Alibaba Shipped Five Qwen Audio 3.1 Models and Cut Audio Prices Up to 95 Percent

Alibaba. Qwen-Audio-3.1 covers speech recognition, text-to-speech, and real-time interaction across five models. The ASR model improves multilingual and dialect handling and strips filler words and repetitions automatically, while ASR-Next adds multi-speaker identification with timestamps and detects emotion, ambient sound, and machine noise. The price cut of up to 95 percent is the part that changes build-versus-buy math for transcription pipelines. Source

2. Microsoft Is Reportedly Authorizing 30 to 50 Percent Copilot Discounts

Microsoft. The Information reports Microsoft told sales staff this week it will approve discounts of 30 percent for customers buying 1,000 to 10,000 Copilot seats and 50 percent above 10,000, in exchange for commitments to usage-based charges on specific features. The change is said to take effect in October alongside a revamped Copilot “super app.” Against a $30 per user per month list price and 20 million paid seats, this is a deliberate shift of revenue from subscription to consumption. Source

3. Enveda Raised $311 Million at a $2 Billion Valuation for Nature-Derived Drugs

Enveda. The AI biotech is running clinical trials on drugs for skin conditions and on preserving weight loss after patients stop GLP-1s, and the new round funds pushing more nature-derived candidates into trials. Enveda’s approach is machine learning over natural product chemistry rather than de novo molecular design. Source

4. Ema Raised $77 Million as Agents Start Displacing Services Budgets

Ema. The round brings Ema to $140 million raised across more than 50 enterprise customers including Google and Microsoft. TechCrunch frames the thesis as AI eating into both enterprise software seats and the services spend layered on top of them, which is a different sales motion from selling a tool. Source

5. Basecamp Research Is Training on Genetic Material Collected From Rainforests and Hot Springs

Basecamp Research. The London company has raised $140 million from investors including NVIDIA and Anthropic’s Anthology Fund, and trains models on genetic data gathered from field expeditions to design antibiotics and cell therapy tools. CTO Philip Lorenz argues biology is a harder problem for AI than language, and that benchmark scores on protein tasks do not predict whether a designed molecule works. Source

6. Nscale Left ByteDance Out of the Main Body of Its IPO Prospectus

Nscale. The NVIDIA-backed AI cloud provider omitted its largest customer, ByteDance, from the main prospectus for its planned US listing. Nscale’s filing already disclosed 85 percent customer concentration, and the omission puts a name to the risk investors are being asked to price. Source

7. Daily AI Users Are Not Less Worried About AI Than Everyone Else

Research. A new survey finds Americans who use AI every day report levels of concern comparable to lighter users, and that exposure does not reduce support for regulation. The finding cuts against the familiarity argument that vendors have leaned on when discussing public opinion. Source

8. A US Plan for AI Safety Alerts Left Out the Technical Experts

Policy. Experts told Ars Technica that framing AI as a race to win against China may deter Beijing from sharing safety intelligence, and that the US proposal for an AI safety alert mechanism omits the technical community that would have to operate it. China has not publicly responded. Source

9. XPRIZE Wildfire Teams Detected Fires in Under 10 Minutes but Could Not Suppress Them

XPRIZE. The $11 million competition produced winning systems that spotted ignitions within 10 minutes, while the suppression half of the challenge went unmet. The split result is a clean illustration of where AI perception currently outruns the physical actuation built around it. Source

10. Spotify Is Letting US Listeners Edit the Model’s View of Their Taste

Spotify. Taste Profile is rolling out to US Premium subscribers, showing how the recommender understands a listener and accepting natural-language corrections to reshape it. Exposing and making a recommendation profile editable is a notably different transparency posture from explaining recommendations after the fact. Source

11. Over 1,000 Developers Told GitHub They Want Tools for Measuring Wasted Compute

GitHub. Research from GitHub and the Yale Program on Climate Change Communication found strong demand among developers for measurement tooling and practical guidance on reducing wasted compute, rather than awareness campaigns. The gap the respondents describe is the absence of a default way to see what a workload actually costs to run. Source

12. Greece’s Prime Minister Said No Government Is Ready for What AI Will Do

Policy. In an interview with TechCrunch, Kyriakos Mitsotakis said governments are “already fighting yesterday’s battle” on AI and conceded that no administration is prepared for the pace of change. The candor is unusual for a head of government on a trade mission. Source