Three Models, One System
In the early hours of July 21 Beijing time, OpenAI held a launch event introducing the GPT 5.6 family. Rather than a single model, the company released three tiers: Soul, Terra, and Luna, each targeting different use cases.
Alongside the models, OpenAI launched ChatGPT Work (a task entry point), a new ChatGPT desktop app (environment access), and Hosted Sites (result delivery).
Taken together, these moves reveal OpenAI's strategy: transform ChatGPT from a conversational tool into an execution system that can accept tasks, read files, call tools, interact with the desktop, and deliver results.
Soul / Terra / Luna Division of Labor
The three GPT 5.6 models have distinct roles:
- Soul: Handles complex Agent workflows, long-context reasoning, and multi-step planning
- Terra: Daily tasks, conversation, and lightweight analysis
- Luna: High-frequency low-cost tasks — summarization, classification, format conversion
The key insight is routing. In Agent scenarios, a complex task may involve reading long documents, multiple tool calls, code generation, and result validation. Using the most powerful model for every step would be too expensive and slow. Using a cheap model for every step would be unreliable for critical operations.
OpenAI's solution is tiered routing. Simple tasks go to Luna and Terra; complex reasoning and code generation go through Soul; intermediate steps switch dynamically. Users see a single ChatGPT, while behind the scenes the system repeatedly selects models, allocates context, and routes tools.
ChatGPT Work: From Q&A to Tasks
ChatGPT Work changes the interaction paradigm. Instead of plain text prompts, users submit tasks. The system determines which data sources are needed, what to read first, what to process later, which steps can run automatically, and where user confirmation is required.
Demonstrated examples include: finding internal use cases from Slack and employee feedback, scheduling interviews, reading financial data for variance analysis, updating forecasts, and generating reports with shareable pages.
The hard part is orchestration and state management. Ordinary Q&A only handles the current input; Agent workflows must track which materials have been used, what conclusions have been drawn, and which assumptions remain unverified.
Desktop App: Real Environment Access
The web version relies on file uploads and cloud connectors. The desktop app can directly access local files, browser tabs, and other applications.
Demonstrated capabilities include organizing Apple Notes, reading local folders, analyzing Chrome tabs, and creating visualizations in spreadsheets.
But environment access brings its own challenges. The model can't stuff every file on a user's computer into its context window. It must first determine what's relevant, then convert PDFs, Excel sheets, web pages, and notes into processable representations. This involves file indexing, semantic retrieval, context compression, citation tracking, and conflict resolution — when PDFs, PPTs, Slack messages, and Excel data contradict each other, the Agent must judge which sources are newer and more reliable.
Hosted Sites: Results Go Live
In the past, LLM output was mostly text or code snippets that users had to move to other tools. Hosted Sites turns results directly into web pages, dashboards, internal tools, or interactive prototypes — suitable for temporary dashboards, project reports, and lightweight scenarios.
The delivery layer needs structured protocols. When an Agent calls a tool, what are the input fields? How are results represented? How are failures handled? Which operations are read-only, and which have side effects? For example, "update the forecast" can't just let the model modify Excel directly. Instead, it should be broken into: read cells, check formulas, generate a diff, wait for confirmation, then write back.
Ultra Mode and Multi-Agent Parallelism
When a single model isn't enough, Ultra Mode splits complex tasks across multiple Agents running in parallel. One Agent reads materials, another processes spreadsheets, another generates pages, another checks consistency — then results are integrated.
The hardest component is the orchestrator. It decides how to split the task, what context each Agent gets, how to handle conflicting results, when to stop, how to avoid redundant work, and how to control costs. Without a stable orchestration layer, multi-Agent systems just make error chains longer.
Security and Trust Boundaries
The closer Agents get to real work environments, the more concrete security concerns become. Traditional model safety focused on output content. In the Agent era, risk falls on actions — which files were read, which spreadsheets were modified, which tools were called, whether messages were sent.
OpenAI noted GPT 5.6's cyber capabilities have improved significantly — it can find vulnerabilities and generate patches. While valuable for defense, this capability also demands stricter permission design for general Agent products.
OpenAI's four-layer security approach:
- Least privilege: Only expose what the task requires
- Operation tiers: Different confirmation logic for reading, drafting, modifying, and sending
- Process audit: Log what the model read, changed, and its reasoning
- Transactional execution: Generate plans and diffs for user review before committing, with rollback support
New Benchmarks
OpenAI mentioned several new evaluation benchmarks: Terminal Bench (code execution), BrowseComp (complex information retrieval), and Agent's Last Exam (long-cycle professional tasks). These better reflect Agent scenarios than traditional Q&A tests.
For real enterprise environments, more relevant metrics include failure rate, human handoff rate, average completion time, tool call success rate, and cost per task.
Competitive Landscape
Anthropic's Claude Fable 5 has been making significant Agent-scene updates, emphasizing code capabilities and long context. The two companies are now competing head-on in the Agent direction. GPT 5.6 differentiates with its three-tier model routing and Ultra Mode multi-Agent orchestration, while Fable 5 focuses on code execution quality.
Conclusion
GPT 5.6's technical roadmap is clear: ChatGPT Work accepts tasks, the desktop app accesses the environment, Hosted Sites delivers results; Soul/Terra/Luna plus Ultra Mode handle model routing; permission and audit mechanisms control risk.
Taken as a whole, OpenAI isn't just releasing a stronger chat model — they're building an Agent operating system. Its ultimate value won't be determined by launch demos, but by real-world stability.