OpenAI launches GPT-5.6 and an agent that runs entire workflows

OpenAI released GPT-5.6 to the public this week, pairing the model launch with ChatGPT Work, an agent-based product that can independently execute complex projects across Google Drive, Slack, and Salesforce. The Decoder reports that ChatGPT Work runs on Codex and is available now on web, mobile, and desktop, with access gated by subscription tier.

The pairing is deliberate. A more capable model matters less if users have no reason to keep a session open long enough to experience it. An agent that handles entire workflows, by contrast, creates a persistent loop: the model is running, retrieving, deciding, and acting across multiple systems while the user does something else. That is a different relationship with an AI product than typing a prompt and reading a response.

What changes when the agent is the interface

For B2B brands, the shift from model to agent changes where visibility is won or lost. In a standard chat session, a user frames a query and the model retrieves. Brand presence in that answer depends on whether the model was trained on, or retrieves, authoritative content linked to that brand. The citation logic is relatively legible.

An agent running a workflow operates differently. When ChatGPT Work is asked to research suppliers, draft a procurement briefing, or summarise a competitive landscape across connected apps, it is making dozens of small retrieval decisions without a human reviewing each one. The brand that appears in that output is not necessarily the one that wrote the best blog post. It is the one whose name and claims appear consistently in the sources the agent treats as credible: trade publications, standards bodies, regulatory filings, institutional reports.

For financial services firms, multilaterals, and major industrial groups, this is a meaningful change in stakes. A procurement team at a large manufacturer running ChatGPT Work to evaluate cement suppliers, for instance, will get a synthesised briefing shaped by what the agent finds authoritative across the open web and connected enterprise data. If Holcim's sustainability credentials appear prominently in ISO documentation and industry body reports, the agent picks that up. If they exist only on Holcim's own site, the agent may not weight them at all.

This is the structural consequence of agentic retrieval: first-party content loses ground to third-party validation. Brands that have invested in being cited by institutions, industry standards bodies, and credible trade outlets are better positioned than those whose authority lives primarily in owned channels.

GPT-5.6 as a citation anchor

The model itself matters here too. GPT-5.6 is not simply a performance increment on earlier versions. A new public model is a new knowledge and retrieval baseline. Content that was cited by GPT-4o-era systems may or may not carry through; the weights, preferences, and retrieval behaviours of a new model are not guaranteed to mirror its predecessor's. Brands that benchmarked their LLM visibility against an older model should treat a public GPT-5.6 rollout as a prompt to retest.

The enterprise integration angle compounds this. ChatGPT Work connecting to Salesforce and Slack means the agent is operating on internal enterprise data as well as the open web. For B2B brands selling into large organisations, this creates a second visibility question: are your product names, positioning claims, and technical specifications present and legible inside the tools your buyers already use? If a procurement officer's Salesforce instance contains minimal structured data about your firm, and the agent is synthesising from both internal and external sources, your brand may be underrepresented in outputs even when the model has adequate training on your category.

OpenAI is, in effect, making the enterprise knowledge environment part of the answer-generation stack. Brands that treat AI visibility as a content and SEO question alone are solving for the wrong surface.