OpenAI Unveils GPT-5.5-Cyber and Three Voice Models

OpenAI has announced a suite of new AI products, including a cybersecurity-focused model and three voice-based systems designed for real-time interaction. The announcement, reported by ForkLog, marks a significant expansion of the company's capabilities in both security and conversational AI.

The centerpiece of the security offering is GPT-5.5-Cyber, a specialized variant of the company's flagship model. Unlike general-purpose AI systems, this neural network has been trained specifically for security workflows: identifying vulnerabilities, classifying threats, verifying patches, and analyzing malware. The model is currently available only to a limited group of partner developers.

OpenAI's blog states that GPT-5.5-Cyber will allow partners to explore complex workflows where access specificity matters. This restricted rollout mirrors the approach taken by Anthropic with its Mythos model, which launched a month prior. Anthropic withheld Mythos from public release due to security risks, instead launching Project Glasswing to test the tool in controlled environments.

The timing is notable. In April, OpenAI released GPT-5.4-Cyber as part of its Trusted Access for Cyber program, which launched in February. The new GPT-5.5-Cyber represents an iteration on that foundation, suggesting the company is treating cybersecurity AI as a product category requiring careful calibration rather than a one-off release.

On the voice front, OpenAI introduced three models simultaneously through its Realtime API:

GPT-Realtime-2 brings GPT-5-level reasoning to voice agents. The model handles complex queries and natural conversation across five reasoning levels: minimal, low, medium, high, and very high. Its context window expanded from 32,000 to 128,000 tokens, which matters because voice agents with persistent memory across long conversations are still rare in practice (a gap that made earlier voice AI feel frustratingly forgetful).

GPT-Realtime-Translate handles simultaneous speech-to-speech translation, keeping pace with the speaker's speech. It translates live speech from over 70 languages into 13 target languages, with a 16,000-token context window.

GPT-Realtime-Whisper is a streaming system for real-time speech-to-text conversion. It can create subtitles during meetings, conferences, lessons, and broadcasts, while also generating notes and summaries during or immediately after conversations.

According to Nicholas Rhodes' AI Brief, the Realtime API simultaneously exited beta and became generally available. This is developer-speak for "stop treating this like a demo" — companies can now ship production products on this infrastructure rather than proofs of concept.

The voice market timing is strategic. ElevenLabs crossed $500 million in annual recurring revenue in early 2026, with customers including Deutsche Telekom, Revolut, Klarna, Meta, and Salesforce. OpenAI's simultaneous launch of three voice models and API graduation suggests the company is treating voice AI as a category worth claiming.

Beyond the core announcements, OpenAI released a Codex extension for Chrome on macOS and Windows. The tool interacts with applications and websites, operating in parallel across all tabs in the background. According to the company's announcement, Codex can handle routine browser tasks like navigating structured pages and performing complex data entry processes. At each stage of task execution, the model independently selects the most suitable tool.

OpenAI also added an "Emergency Contact" feature to ChatGPT. Users can designate a contact person for emergencies. If the company's trained specialists detect that a user has been asking questions about self-harm, the trusted contact receives a notification via messenger, email, or another ChatGPT account. This follows parental controls implemented in September 2025.

The cybersecurity angle carries regulatory weight. The White House recently briefed executives from Anthropic, Google, and OpenAI on plans for a working group to vet AI models before public release. Officials walked it back within days, calling the reports speculation. The catalyst appears to be Anthropic's Mythos model, which can map software security vulnerabilities at scale — a capability that could compress the window between zero-day discovery and exploitation down to hours.

Whether developers actually adopt these voice models at scale remains the real question. The infrastructure is now production-ready, but the market will decide if voice agents can move beyond novelty into genuine utility.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn