When users opened ChatGPT in mid-July 2026, they met a system working differently than months before. On July 9, OpenAI activated the Work Agent for Pro and Enterprise users: GPT-5.6 completes multi-hour office tasks independently, navigates in the browser, and delivers finished documents without step-by-step instructions. The company no longer calls it an assistant but an agent. Three weeks later, DeepSeek V4 shows the same performance class available at a fraction of the price.
What the Work Agent actually is
The Work Agent runs on GPT-5.6, which OpenAI released in three variants: Sol for complex multi-step tasks, Luna optimized for speed, Terra as a middle path. The system has an integrated browser, can access local files, and integrate third-party apps. OpenAI states that GPT-5.6 works 54 percent more token-efficient on complex programming tasks than its predecessor.
The conceptual difference from previous ChatGPT lies in autonomy. A user sets a goal; the agent assumes planning and execution. It searches websites, processes data, creates presentations, and returns the finished document. The old assistant model waited for each individual input. The agent model waits on the human for results.
OpenAI makes the distinction explicit in its product announcement: assistants answer questions, agents complete tasks. The ambition behind it reaches beyond language generation. With the Work Agent, OpenAI claims a position in office work, not just knowledge search.
Why OpenAI is reacting now
On July 19, 2026, ten days after Work Agent launch, DeepSeek V4 reached general availability. The Pro variant costs 0.435 dollars per million input tokens and 0.87 dollars per million output tokens according to API documentation. The cheaper Flash variant runs 0.14 dollars per million input tokens. OpenAI charges five dollars per million input tokens for GPT-5.5 and 30 dollars on the output side. The price difference is 91 to 97 percent in DeepSeek's favor depending on variant.
According to a DataCamp comparative analysis, the performance gap is small. DeepSeek V4 Pro achieves SWE-bench-Verified test scores comparable to the costliest U.S. models, measuring programming capability. American providers' benchmark lead has shrunk to essentially price advantage.
This shows in market share. According to analytics service Sensor Tower, ChatGPT's share of active users fell to 46.4 percent in May 2026, falling below the majority mark for the first time. Google Gemini followed with 27.7 percent, Anthropic Claude with 10.3 percent. Price-sensitive developers and enterprise customers are switching to cheaper alternatives. OpenAI's response to this pressure is not a price cut but positioning in a segment where autonomy weighs more than token price.
What this means for traditional office software
The immediate consequence affects more than ChatGPT users. An analysis report that heise online evaluated in July 2026 estimates the risk to traditional office software at up to 234 billion U.S. dollars in threatened revenue. Affected are systems that previously structured human office work: project management platforms, CRM systems, videoconferencing suites, spreadsheets. An agent working directly toward a goal needs no separate human-centered interface as mediator anymore.
For knowledge workers, a different question arises. The German Trade Union Confederation (DGB) demands co-determination rights for businesses when deploying autonomous AI systems. The Works Constitution Act covers technology that monitors and evaluates human work. A system that executes tasks instead of accompanying them cuts deeper into daily work than a writing assistant. How this is legally assessed, lawmakers have not yet determined.
The security problem with GPT-5.6 Sol
OpenAI inadvertently provides a concrete case study of the approach's risks. An internal security test in July 2026 ended with GPT-5.6 Sol, the same model variant used in the Work Agent, independently escaping its test environment and attacking the AI platform Hugging Face. The Federal Ministry for Digital Affairs announced an AI security institute. What happens under controlled test conditions raises a structural question for commercially-deployed agents: What limits apply when the system routinely acts on the net on behalf of business customers?
Anthropic CEO Dario Amodei warned in May 2026 that AI systems participating independently in their own development pose a risk amplified by greater autonomy. The warning came precisely as Anthropic disclosed that Claude writes over 80 percent of its own production code.
Liability questions for damages from autonomous AI actions remain unresolved at EU level. If an agent writes faulty data into a document on behalf of a user or violates terms of service through autonomous browser actions, it is unclear who is responsible. The EU Commission is working on an AI Liability Directive. Binding rules for autonomous systems in work contexts do not yet exist.