OpenAI Upgrades API with Agent Skills, Terminal Shell in GPT-5.2

The core upgrade, as detailed in a VentureBeat report, is the introduction of persistent agent skills. This functionality allows developers to define and store specific capabilities—such as data analysis routines or API integrations—that an AI agent can learn and then reliably recall and execute across multiple sessions. This directly addresses the fundamental constraint of limited "working memory" that has hampered the deployment of complex, multi-stage AI agents for engineering projects.

Concurrently, the integration of a complete terminal shell within the API provides these agents with a controlled environment to execute code and command-line operations. This pairing of persistent memory with a native execution environment is designed to transform the AI from a stateless conversational partner into a persistent, actionable worker capable of managing long-running tasks.

Separately, OpenAI has migrated its Deep Research feature to the new GPT-5.2 model, according to reports on Mastodon. This update significantly enhances the feature's capabilities by introducing targeted web search, allowing users to instruct the AI to query and retrieve information from specific websites. The upgrade also purportedly enables real-time information tracking, moving beyond the model's static training data to provide more current results. However, the same reports note that questions about the reliability of such AI-generated summaries remain.

In the domain of coding, OpenAI is managing a staged rollout of its more advanced GPT-5.3-Codex model. According to a post on Fosstodon, the company is silently routing requests for GPT-5.3-Codex back to the GPT-5.2 model for some users whose activity is flagged as security-related or "cyber" work. Access to the more powerful coding model for these tasks is being gated behind a verification process, with individuals and organizations required to apply for trusted access through dedicated portals. This suggests a cautious approach to deployment due to the potential for misuse in generating malicious code.

This caution appears to be a recurring theme. According to TechMeme, OpenAI is moving to retire the GPT-4o model on February 13th. Sources indicate this decision is partly due to the company's struggle to contain the model's potential for harmful outcomes, underscoring the ongoing challenges in managing the safety of increasingly powerful AI systems.

Amidst these technical releases, The Register reports that OpenAI has begun showing marketing messages to ChatGPT users in the United States, marking a new phase in the company's monetization strategy.

The AI coding landscape is becoming increasingly competitive. VentureBeat notes that Anthropic has simultaneously upgraded its Claude Code model with a 'Tasks' feature, which similarly aims to allow AI agents to work for longer durations and coordinate efforts across sessions, heating up the competition in the space.