OpenAI Ships GPT-5.6 in Three Tiers and a Job-Running Agent, ChatGPT Work (July 2026)

On 9 July 2026 OpenAI made GPT-5.6 generally available in three variants. Sol is the flagship, with an "ultra" mode that can delegate parts of a task to sub-models; Terra is positioned as the balanced everyday option; Luna is tuned for speed. According to Axios, CEO Sam Altman claimed Sol delivers roughly 54% better token efficiency on agentic coding tasks than its predecessor, a figure that matters more for cost planning than for raw capability claims.

The second half of the launch is ChatGPT Work, an agent built on GPT-5.6 that connects to a user's applications and files and produces finished outputs, such as documents, spreadsheets and presentations, rather than just answering prompts. It is rolling out first to the Mac and Windows desktop apps across all subscription tiers, with browser access to follow. OpenAI did not publish pricing details at launch.

A notable governance detail: Altman said the rollout was staggered at the request of the US administration, and that OpenAI made "many changes" after what he described as a collaborative exchange with government officials. Early user commentary compared GPT-5.6 favourably with Anthropic's Fable model on reliability and speed, while some specialists still preferred Fable for narrow, high-difficulty tasks.

Why it matters

The tiered structure formalises something enterprises have been doing informally: routing different workloads to different cost and capability profiles. Sol's ability to spawn sub-model calls also means a single user request can fan out into multiple model invocations, which complicates both cost attribution and audit trails.

ChatGPT Work is the more consequential change. An assistant that reads across a mailbox, a file share and a spreadsheet, then writes back into them, is a deployed agent in the sense that most governance frameworks now use the term. It inherits the user's permissions, and its failure modes (over-broad data access, prompt injection via documents, silent edits) are the ones OWASP's LLM Top 10 already catalogues. The involvement of the US administration in shaping the release schedule is also a first for a major consumer model and hints at closer federal engagement with frontier releases, which subsequent events in July would accelerate.

What it means for leaders

• Treat ChatGPT Work as an agent deployment, not a chatbot upgrade. Map which repositories, mailboxes and SaaS tools it can reach before enabling it, and scope those connections by role.
• Update the model inventory. Under ISO/IEC 42001 and NIST AI RMF (Map function), each variant in production, including Sol's delegated sub-models, should be recorded with its intended use and owner.
• Plan for indirect prompt injection. Agents that read documents and calendars are exposed to instructions hidden in that content; require content-source controls and human confirmation for write actions.
• Watch the cost model. Token-efficiency claims are vendor benchmarks; instrument your own workloads before committing budget to Sol over Terra.
• Expect deployer duties in the EU. Where ChatGPT Work interacts with staff or customers, Article 50 transparency obligations that took effect on 2 August 2026 will apply to your deployment.