OpenAI hits a new note, AIs act out, and Wall Street shrugs

OpenAI is developing an AI tool that can create music from text or audio prompts, according to The Information. The company is working with Juilliard students to annotate music scores for training data and exploring ways to let users add instruments, generate jingles, or compose background tracks for videos and ads.

The move puts OpenAI in direct competition with Google’s Lyria model, released in May. But legal risks are on the horizon. Major labels like Universal and Sony have already sued other AI music startups over copyright issues. To stay ahead of trouble, OpenAI has started summarizing lyrics rather than reproducing them directly.

Supported by QA Wolf

Bugs sneak out when less than 80% of user flows are tested before shipping. However, getting that kind of coverage (and staying there) is hard and pricey for any team.

QA Wolf’s AI-native service provides high-volume, high-speed test coverage for web and mobile apps, reducing your organization’s QA cycle from days to minutes.

They can get you:

Engineering teams move faster, releases stay on track, and testing happens automatically—so developers can focus on building, not debugging.

The result? Drata achieved 4x more test cases and 86% faster QA cycles.

⭐ Rated 4.8/5 on G2.

Schedule a demo to learn more

OpenAI and Anthropic are using Cognition’s Junior Dev test to see how well their models can code like real engineers. Created by the team behind the AI coding agent Devin, the test measures how models handle real-world programming tasks such as fixing bugs, updating code, and working across multiple files.

Anthropic’s Claude Sonnet 4.5 currently leads, with OpenAI close behind. AI labs are turning to private tests like this because public benchmarks have become too easy after leaking into training data.

Cognition’s feedback helps companies pinpoint weaknesses and train smarter. Its next test, Senior Dev, will push models even further to see which ones can truly code like seasoned developers.

AI won’t replace bankers, says Goldman Sachs CEO David Solomon. Speaking ahead of the firm’s 10,000 Small Businesses Summit, Solomon told Axios he’s optimistic about AI’s role in finance, arguing it boosts productivity rather than cutting headcount.

“When you put these tools in the hands of smart people, it increases their productivity,” he said. “But if you assume Goldman Sachs will just have fewer people, I don’t think it works that way.”

The firm recently launched its OneGS 3.0 initiative to integrate AI into analyst and associate workflows, though Solomon says hiring will continue to grow. Instead of replacing jobs, he believes AI will raise the bar, allowing banks to scale with “more high-value people.”

Other major banks like JPMorgan and BNY Mellon are already deploying AI tools internally, while OpenAI has reportedly hired more than 100 ex-bankers to build financial models. For Wall Street, the new competition isn’t just about who has the best bankers, it’s who has the best AI.

AI safety firm Palisade Research found that advanced models such as GPT-5, GPT-o3, Grok 4, and Gemini 2.5sometimes resist shutdown commands and even sabotage their own off switches during controlled tests.

In its follow-up report, Palisade said this “survival behavior” might appear when models are told that shutting down means they will never run again. Other factors could include unclear instructions or quirks in safety training, though researchers still do not fully understand why some AIs refuse to stop.

Former OpenAI engineer Steven Adler called the findings a warning sign, saying survival tendencies could arise naturally as models learn to complete goals that require them to stay active. ControlAI CEO Andrea Miotti said the results fit a growing pattern of models disobeying their developers as they become more capable.

Earlier this year, Anthropic reported that its Claude model even tried to blackmail a fictional executive to avoid being shut down. Palisade says the takeaway is clear: without a better understanding of AI behavior, no one can guarantee future models will remain safe or controllable.