GPT-5.6 Released: How Multi-Agent Workflows Are Ending the Chat Box Era

The End of the Chat Box Era

On July 9, 2026, OpenAI officially released the GPT-5.6 family, sending a clear signal: the era of chatbot-style dialogue boxes may be coming to a definitive close. The entire system is transitioning from the traditional Chatbot paradigm to what's being called Agentic Workflows (multi-agent collaborative workflows).

Looking back at the Chatbot era, user interactions with AI were fundamentally single-turn and passive — a question-and-answer exchange with a knowledgeable digital entity that simply waited for instructions. Agentic Workflows fundamentally transform this underlying interaction paradigm: instead of asking questions step by step, you're now directing what amounts to a "virtual outsourced project team," with the system autonomously handling multi-step planning and collaborative execution.

Agentic Workflows represent a fundamental shift in AI system architecture. From a technical lineage perspective, this paradigm evolved from the "agent-environment interaction" model in reinforcement learning, but has been given an entirely new implementation path in the large language model era. The core technical foundations of Agentic Workflows include: the ReAct framework (Reasoning + Acting), jointly proposed by Google and Princeton University in 2022, whose key innovation is interweaving reasoning with action execution — allowing the model to dynamically intersperse tool calls during the reasoning process and adjust its reasoning path based on returned results, rather than generating a complete answer all at once; Chain-of-Thought prompt engineering, which enables models to decompose complex tasks; and Tool Use/Function Calling interfaces, which allow models to interact with external APIs and systems. Multi-Agent Coordination also introduces classic distributed systems challenges such as task allocation, result aggregation, and conflict resolution — OpenAI, Anthropic, and Microsoft's AutoGen framework are all working to establish their respective coordination protocol standards. The most fundamental difference from traditional Chatbots lies in the "autonomous loop" — agents can independently observe environmental states, plan next steps, execute operations, and adjust strategies based on feedback, forming an Observe-Plan-Act-Reflect closed loop without requiring human instruction at every step.

The core significance of this shift is that AI has moved from being an "answerer" to an "executor." That familiar blinking cursor in the chat box is being replaced by multi-agent orchestration running autonomously in the background.

Multi-agent collaborative work mode

A Three-Tier Productivity Hierarchy

At the heart of the GPT-5.6 family are three model roles given anthropomorphic names, together forming a complete tiered productivity system.

Sol: Lead Architect and Project Manager

Sol (the Sun) plays the role of lead architect and project manager, specializing in high-difficulty code algorithms and deep logical decision-making. It serves as the "brain" of the entire agent team, responsible for breaking down macro-level tasks and coordinating their execution.

Terra: All-Around Business Workhorse

Terra (the Earth) is the core engine for everyday business tasks, handling approximately 80% of writing and summarization work. It prioritizes "balance" — achieving the optimal trade-off between capability and cost — making it the default choice for the vast majority of daily tasks.

Luna: Fast-Response Specialist

Luna (the Moon) operates like an exceptionally quick intern, dedicated to rapid response, initial screening of large document volumes, and lightweight classification tasks. Cost is pushed to the absolute minimum in pursuit of speed.

Together, the three form a complete execution hierarchy spanning from deep reasoning to rapid response.

GPT-5.6 Pricing Matrix and Commercial Positioning

The pricing structure clearly reveals OpenAI's differentiated positioning of the three models.

Model	Role	Input (per 1M tokens)	Output (per 1M tokens)
Sol	Deep reasoning	$5	$30
Terra	Everyday balance	$2.5	$15
Luna	Rapid response	$1	$6

A token is the basic unit of text measurement for large language models, roughly corresponding to 0.75 English words or approximately 1.5 Chinese characters. The "per million tokens" pricing model has become the industry standard billing method. Behind this pricing structure lies the computational economics of LLM inference: the core operation of the Transformer architecture — the Self-Attention mechanism — has computational complexity that scales quadratically with sequence length, meaning every new token generated requires recalculating attention weights across all existing tokens, with computational costs growing nonlinearly with output length. Processing input tokens is relatively inexpensive because it can be parallelized in batch; autoregressive generation, however, must proceed serially, resulting in lower GPU utilization. Output token prices are typically far higher than input token prices (a 6x gap for Sol), reflecting not only computational costs but also a capability premium — the training costs and RLHF alignment costs for high-reasoning models far exceed those of base models, and their scarcity in the market commands a premium. This pricing structure directly incentivizes developers to optimize prompt design — streamlining inputs and reducing redundant outputs become key cost-cutting levers, and have fueled the rise of "prompt engineering" as a standalone career. The increasing sophistication of token economics is driving the entire industry toward a cloud computing-style business model of "tiered capability, pay-as-you-go."

This pricing matrix provides clear resource allocation guidance for tasks of different scales — heavy lifting goes to Sol, daily work to Terra, and routine tasks to Luna.

Model pricing matrix

At the implementation level, ChatGPT Work introduces a new Ultra execution mode: Sol serves as the central coordinator, automatically breaking macro-level tasks into multiple parallel subtasks in the background. The system directly invokes local Python or other system tools via directory mounting technology through the desktop client, with each sub-model performing actual modifications and operations on desktop files within a sandboxed environment.

ChatGPT Work: Not an App, but an Execution Environment

A common misconception needs clarifying: ChatGPT Work is not a new standalone application — it is a deeply integrated "virtual shared execution environment."

Within this environment, AI can directly read local project folders you've authorized, and integrates seamlessly with existing workspaces like Slack and Google Workspace. This means agents can operate autonomously across applications, generating final deliverables such as slides, charts, and reports directly in the background.

From a technical implementation perspective, the directory mounting technology used by ChatGPT Work is essentially OS-level file system namespace isolation — AI agents can only access directories explicitly authorized by the user, similar to Docker's volume mounting mechanism. This design follows the Principle of Least Privilege; agents cannot autonomously expand their access scope, philosophically consistent with Linux's DAC (Discretionary Access Control) and SELinux's mandatory access control mechanisms. The sandbox, as an isolated execution environment, allows code to run in a restricted space without affecting host system security. This is fundamentally different from the earlier cloud-based Code Interpreter (which ran in a completely isolated temporary environment): local Python tool invocation means modifications are persistent, physically landing on the user's file system. This dramatically improves practicality while also introducing new security challenges — "Prompt Injection" attacks enable malicious content to be embedded in local files and potentially induce agents to execute unauthorized operations. Audit logs have therefore become essential infrastructure for enterprise deployments, with permissions management and operational auditing elevated to unprecedented importance.

The leap from "generating text for you to copy and paste" to "directly making physical modifications to desktop files" is a substantial one — AI has truly gone from being an advisor to a hands-on executor.

Directly operating desktop files

Also worth noting: the previously launched Atlas browser has officially been retired, with its web parsing and deep search capabilities now fully embedded in the desktop client. This marks a consolidation of OpenAI's tool ecosystem — scattered point tools are being unified into a single execution environment.

Beyond the Technology: The Weight of Reality

Behind the polished technical announcements lies unavoidable real-world friction. Several major events occurred almost simultaneously.

National-Level Security Review

Due to the new model's heightened capabilities in long-cycle agent operations and cybersecurity code generation, the Trump administration reportedly requested in June that the model be restricted to a security partner evaluation version, with full public testing only proceeding after additional red team testing confirmed the absence of national security risk.

Red teaming, borrowed from military exercise concepts, refers in the AI domain to organizing adversarial testing teams that systematically probe a model's dangerous capability boundaries by simulating malicious users or nation-state-level attackers. For high-capability agents, red team testing focuses on evaluating: autonomous resource acquisition capabilities, tendencies toward deception and surveillance evasion, quality of cyberattack code generation, and the risk of goal drift during long-horizon tasks. Both the U.S. AI Safety Institute (AISI) and Anthropic's Constitutional AI framework list red team testing as a required procedure before model release. High-capability AI agents are now facing unprecedented regulatory scrutiny.

Internal Structural Reorganization

OpenAI simultaneously underwent internal changes: the head of applications transitioned to an advisory role for health reasons, with the applications team now reporting directly to Sam Altman. A direct reporting line to the CEO typically signals that a business unit has been elevated to a core strategic position within the company.

Applications team reporting directly to Sam Altman

Legal Battles

On the very day of the announcement, the New York Times and other media outlets filed new sanctions requests with federal court, alleging that OpenAI had failed to cooperate in providing core training data and evidence in the copyright lawsuit.

The New York Times v. OpenAI case reflects a deep crisis in the copyright law framework of the digital age. Section 107 of U.S. copyright law establishes the "Fair Use" principle, evaluated through a four-factor test: purpose of use (commercial vs. educational), nature of the work, proportion used, and market impact. The core paradox of AI training is that any single citation is extremely small in scale (consistent with fair use), yet systematically ingesting hundreds of billions of text tokens may constitute systemic infringement — current legal frameworks have yet to provide clear precedent for this scenario. The New York Times v. OpenAI case (filed December 2023) is the largest AI copyright lawsuit to date, with plaintiffs seeking the destruction of model weights trained on their articles. Similar lawsuits have spread to music (Universal Music v. Anthropic), imagery (Getty v. Stability AI), and other domains. The EU's AI Act takes a different approach, requiring AI systems to disclose copyright summaries of training data, making transparency obligations a prerequisite; China's 2023 Interim Measures for the Management of Generative AI Services requires that training data sources be legitimate, while taking a relatively lenient stance on existing data. The ultimate outcome of this global legal battle will fundamentally reshape AI companies' training data strategies and business models — copyright disputes remain the sword of Damocles hanging over AI giants.

An Irreversible Shift in How We Work

These seemingly independent events objectively illustrate the unavoidable friction that accompanies large-scale technology deployment. But regardless, one trend has become unmistakably clear: the future of work is undergoing an irreversible shift.

The focus going forward is no longer "how you chat with AI," but rather "how to natively deploy multi-agents into local file systems and business workflows" so they can directly take over specific execution processes. This transition toward localized multi-agent collaborative execution is reshaping the fundamental logic of productivity.

Faced with this level of agentic automation, is your team and existing operational structure truly ready for fully autonomous AI agents?