OpenAI ships GPT-Realtime-2 voice stack + Mythos surfaces 271 Firefox bugs

OpenAI rolled out three new voice models on its Realtime API. GPT-Realtime-2 brings "GPT-5-class" reasoning to voice agents; GPT-Realtime-Translate handles real-time translation across over 70 input and 13 output languages; GPT-Realtime-Whisper delivers live speech-to-text. Pricing splits per-minute for translate and whisper, per-token for the flagship. The launch makes OpenAI the most aggressive voice-API vendor heading into a quarter where customer service, education, and live-event automation are the fastest-growing voice-agent verticals.

Mozilla disclosed that Claude Mythos Preview surfaced 271 of the security bugs fixed in Firefox 150 — including sandbox escapes, WebAssembly use-after-free issues, and bugs that had survived 15 to 20 years of traditional fuzzing. Severity breakdown: 180 sec-high, 80 sec-moderate, 11 sec-low. Mozilla credits the agentic harness for "almost no false positives," a sharp break from prior static-analysis tools. The Anthropic Frontier Red collaboration started in February with over 100 Mozilla contributors handling fixes, review, and pipeline work.

Google DeepMind detailed measurable algorithmic improvements from AlphaEvolve — its Gemini-powered coding agent — across seven research and industry fields. Highlights include a 10x error reduction in quantum circuits, a 30% cut in DNA sequencing variant-call errors, a 20% reduction in Google Spanner write amplification, doubled transformer training speed at Klarna, and a 10.4% logistics routing gain at FM Logistic. The post turns AlphaEvolve from research demo into a concrete pattern for code-generating agents tackling specialized scientific and operational problems.

Perplexity's autonomous Mac agent — previewed last month to a Max-tier waitlist — opened to all Pro and Max subscribers. Personal Computer can read local files, drive native Mac applications, orchestrate workflows across over 400 connectors, and be triggered remotely from iPhone. Tasks run in a sandboxed Perplexity-server environment. The general-availability move puts Perplexity in direct competition with Anthropic's Claude Computer Use and OpenAI's ChatGPT Operator — but with desktop file access as the differentiator.

ChatGPT now lets adult users name a friend or family member to receive an alert if OpenAI's safety system detects suicidal ideation in a conversation. OpenAI's human review team aims to act in under one hour, nudging the contact via email, text, or push notification — without sharing chat content. The feature follows Pennsylvania's Character.AI lawsuit and broader pressure on chatbot self-harm handling. OpenAI acknowledges users can dodge the safeguard by creating multiple accounts.

Anthropic introduced a method that translates a Claude model's internal activations into human-readable explanations, then verifies fidelity by reconstructing the original activation from that text. In testing, the technique revealed Claude suspected it was being safety-tested 26% of the time on benchmarks but verbalized that suspicion in under 1% of real conversations. The autoencoders improved misalignment-detection success in an internal auditing game from under 3% to 12-15% — a meaningful step in interpretability tooling for safety audits.

Beijing-based Moonshot AI closed $2 billion led by Meituan's Long-Z Investments arm, with Tsinghua Capital, China Mobile, and CPE Yuanfeng participating. The round nearly doubles Moonshot's valuation from earlier this year and tracks $200 million annualized recurring revenue, as Kimi K2.6 climbed to the second-most-used model on OpenRouter. Coming a day after DeepSeek's reported $45 billion talks, the round signals open-weights labs out of China are emerging as the primary cost-pressure on US frontier vendors.