OpenAI Uncovers GPT-6.1 from Internal Test Code: Codenamed Terra, Focused on Precise Computer Operations

Erin, author from tech media Testing Catalog, recently published an article stating that OpenAI has launched a new large model in its internal test environment, with relevant entries buried in internal front-end test code and not made public.

The model’s internal identifier is gpt-6.1-terra, with external version number GPT-6.1 and internal codename Terra. Its core capability is labeled as "precise computer use", referring to accurate computer operations.

For regular users, this capability differs significantly from ChatGPT’s previous usage. The model can directly take over device mouse and keyboard permissions, output operation instructions by recognizing screen content, and autonomously complete cross-software tasks. For example, when organizing quarterly expense reports, it can independently open a browser to download e-invoices, switch to Excel for data entry, and then log in to the OA system to submit approvals—without humans needing to break down steps and guide it one by one.

Testing Catalog is a tech media outlet specializing in tracking OpenAI’s developments, long mining unreleased features by analyzing front-end codes and interface configurations. Earlier internal test news about the GPTs Store, ChatGPT real-time voice mode, and Operator computer agent was first reported by this site, and later matched official launch information quite well.

Computer operation has been the core focus of efforts in the large model agent track over the past year. The Operator agent (OpenAI’s previously small-scale gray-scale test) and Anthropic’s publicly launched Computer Use interface both follow the underlying logic of recognizing screen pixels via multimodal models and generating corresponding operational actions. Currently, industry-wide general solutions are stuck in precision bottlenecks: they easily misclick pop-up ads, misidentify function buttons, deviate from initial goals when tasks exceed 10 steps, and often crash directly when encountering unfamiliar non-standard software interfaces.

Inferred from the "precise computer use" label, Terra’s core optimization direction is to solve these common problems.

As of press time, regular ChatGPT users and API developers cannot access the model, and relevant resources are only available to OpenAI’s internal teams and a small number of partners. The official has not responded to related revelations and has never publicly disclosed R&D plans for the GPT-6 series.

There is a detail deviating from industry general expectations: previously, most external parties speculated OpenAI’s next publicly released large version would be GPT-5, but this time it jumps directly to 6.1, likely for two reasons. First, internal version counting and external product names may not be interconnected, so Terra may change its official name when finally launched; second, computer operation capabilities are defined as the core differentiated selling point of the next-generation model, leading to a separate version line.

According to OpenAI’s past release rhythm, the interval from when front-end tracking points appear to small-scale open testing is usually 1 to 3 months. Of course, following their previous practice of temporarily cutting features and delaying releases, it is not too late to make judgments until it is truly available for use.