OpenAI expanded one interface while narrowing two others: three realtime audio models entered the API, GPT-5.5-Cyber stayed identity-gated, and self-serve fine-tuning began a phased closure. Trial testimony tested governance history, Anthropic translated activations into reconstructive explanations, and Perplexity gave a Mac agent access to local files and hundreds of connectors.
1. OpenAI launches three realtime audio models
OpenAI released GPT-Realtime-2 for voice reasoning and tool use, GPT-Realtime-Translate for speech translation from more than 70 input languages into 13 output languages, and GPT-Realtime-Whisper for streaming transcription. All three became available through the Realtime API with separate token or per-minute pricing.
The launch moves voice from turn-taking toward tool-using workflows, but OpenAI supplies the benchmark gains and customer examples. An application-level test matrix can isolate accented speech, interruptions, domain terms, translation lag, tool-call recovery, and unsafe requests under real traffic instead of inheriting one aggregate launch score.
Sources: Advancing voice intelligence with new API models · OpenAI launches new voice intelligence features
2. Former OpenAI leaders describe governance failures at trial
Former CTO Mira Murati testified that Sam Altman created distrust and at times told different executives conflicting things. Separate testimony described concern that ChatGPT and an early GPT-4 deployment advanced without expected board or safety communication, while witnesses also acknowledged the need for substantial commercial funding.
These statements were testimony in Musk's civil case, not judicial findings. They matter because the case is testing whether OpenAI's nonprofit oversight meaningfully constrained product and financing decisions, but credibility, causation, and the requested structural remedies remained for the court to decide.
Sources: Former OpenAI technology chief testifies about Altman's management · OpenAI's safety record examined during Musk trial
3. Anthropic turns model activations into natural-language explanations
Anthropic introduced Natural Language Autoencoders, which train one model copy to verbalize an activation and another to reconstruct that activation from the explanation. The company used the method in simulated safety tests and released code plus trained examples for several open models.
Anthropic warns that the explanations can invent details, are expensive to generate, and cannot be treated as literal transcripts of model thought. The method is better suited to finding themes for corroboration than to continuous monitoring or making a high-stakes decision from one explanation.
Sources: Natural Language Autoencoders
4. Perplexity opens Personal Computer to paid Mac users
Perplexity made its Personal Computer agent available through a new Mac app to Pro and Max subscribers without the earlier waitlist. The agent can work with local files, native applications, web tools, and more than 400 connectors, with remote task initiation and approvals from an iPhone.
Broad file and application access raises the impact of prompt injection, overbroad permissions, and mistaken actions. Perplexity describes a secure server-side development environment, while local files, connector credentials, and remote phone approvals cross additional boundaries. A connector inventory and destructive-action tests can reveal authority that the product label conceals.
Sources: Perplexity's Personal Computer opens to Mac users · Personal Computer is available to all Mac users
5. GPT-5.5-Cyber enters limited preview under stronger identity controls
OpenAI began a limited GPT-5.5-Cyber preview for verified defenders handling specialized work such as authorized penetration testing and exploit validation. Its broader GPT-5.5 model remains the recommended option for most defensive workflows under Trusted Access for Cyber, with fewer refusals for approved users.
OpenAI explicitly says the cyber variant is more permissive rather than uniformly more capable than GPT-5.5. The deployment design makes verification, approved scope, phishing-resistant authentication, monitoring, and partner feedback core controls for a dual-use model whose outputs can enable exploitation.
Sources: Scaling Trusted Access for Cyber
6. OpenAI starts winding down self-serve fine-tuning
OpenAI's deprecation notice said organizations that had never fine-tuned could no longer create training jobs from May 7. Additional restrictions were scheduled for July 2, with active existing customers set to lose the ability to create new jobs on January 6, 2027.
Inference on existing fine-tuned models remains available until each base model is deprecated, making this a two-stage migration rather than an endpoint removal. First, new training jobs disappear; later, each base model sunsets. Model inventories and retained evaluation sets can carry behavior checks across prompt, retrieval, or replacement-model alternatives.
Sources: OpenAI API deprecations and fine-tuning availability