engblogs

Running Codex safely at OpenAI: deploying bounded, sandboxed execution with approval workflows and agent-aware telemetry to balance developer productivity with enterprise security and auditable governance.

Codex is engineered for safe, enterprise deployment by enforcing bounded execution, sandboxed actions, approval gates, strict network and credential controls, and agent-native telemetry for auditability.

Deploys Codex within strict sandboxed boundaries, uses approval workflows, and leverages agent-aware telemetry and compliance logging to balance rapid development with enterprise security.

A Contextual Sequential Two-Tower Model that injects real-time context into a Transformer-based recommender, trained with synthetic context data and deployed via a hybrid offline-online serving flow to boost Related Pins ad relevance and ROAS.

Introducing a Contextual Sequential Two-Tower model that injects real-time context through a context layer and synthetic data, enabling hybrid offline-online inference to boost ad relevance on Related Pins.

Adaptive Parallel Reasoning (APR) enables models to dynamically balance parallel and sequential reasoning at inference time, using fork-join architectures and KV-cache strategies to improve latency and throughput across engine-modified and engine-agnostic implementations.

Explores adaptive parallel reasoning as a paradigm shift for inference, detailing how models can dynamically balance sequential and parallel execution, orchestrate multiple threads, and manage KV cache to achieve faster, more efficient reasoning with flexible, engine-agnostic design.

Explores how Modular Cloud reimagines LLM inference routing with cache-aware, multi-step orchestration and disaggregated GPU pods to achieve predictable latency in stateful backends.

Simplex rethinks software delivery by deploying Codex with ChatGPT Enterprise to accelerate design, development, and testing with measurable productivity gains.

Overview of Trusted Contact in ChatGPT, a safety feature enabling a trusted adult to be notified during potential self-harm crises to connect users with real-world support while preserving privacy and undergoing human review.

Introducing three real-time voice models—GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper—that enable live reasoning, multilingual translation, and streaming transcription to power end-to-end voice interfaces that listen, understand context, translate in real time, transcribe as you speak, and take action within conversations.

Simplex rethinks software delivery by deploying ChatGPT Enterprise and Codex to validate AI-driven development and accelerate productivity, achieving 70% faster screen development, 40% faster design, and 17% faster internal integration testing through AI-native workflows.