What shipped on 9 July 2026

On 9 July 2026, OpenAI introduced the GPT-5.6 family and ChatGPT Work. The family separates capability profiles into Sol, Terra, and Luna. ChatGPT Work is an agent for multi-step projects that can work across files and connectors and produce finished artifacts.

The practical change is the unit of work. A chat answer is a response; a Work assignment is a project with inputs, intermediate decisions, and a deliverable such as a document, deck, spreadsheet, or Site. This makes briefing and review more important, not less.

Workshop: What shipped on 9 July 2026

The practical objective of this chapter is separating product facts from conclusions about your workflow. Before opening a chat or Work, write down how the job is performed today, who owns it, and what counts as an acceptable result. Choose one example you can personally inspect. This baseline prevents the novelty of the tool from being mistaken for real improvement and gives you a fair comparison for elapsed time, accuracy, corrections, and usefulness.

Worked example: a team lists the new models and Work capabilities, then tests one document assignment instead of assuming every existing process should change. The team first narrows the outcome and assembles this pack: the public launch reference, current plan screens, and a dated record of the team's own test. It then runs the agent with an explicit stopping condition, checks every material claim, and records corrections. The example succeeds only when a named owner can explain why the result was accepted, which parts remained human work, and exactly what should change on the next run.

The most common misapplication is repeating launch language as if it were measured performance in your organization. The correction is not automatically a stronger model. First reduce scope, rank sources, reveal hidden constraints, and add an approval checkpoint. If the problem remains after those changes, test a higher tier or effort setting on the same sample. This allows you to compare cause and effect instead of merely comparing two different-looking outputs.

• Describe the current human process in five sentences.
• Prepare an evidence pack that includes the public launch reference, current plan screens, and a dated record of the team's own test.
• Set one measurable acceptance rule and one prohibited action.
• Ask the reviewer to classify errors by severity, not only style.
• Save the brief, settings, output, and corrections as one run record.