GPT-6.1 Terra advances visual understanding and computer use
What changed
OpenAI released GPT-6.1 Terra on October 26, combining visual understanding and computer use for common professional tasks. It is available through selected ChatGPT Work and Codex subscription plans and the API.
The breakdown
Access for everyday professional tasks
The announcement lists access for Plus, Pro, Enterprise and Edu accounts on ChatGPT Work and Codex. Developers can also call the model through the API using the identifier gpt-6.1-terra. Visual analysis, form filling and video editing are among the example tasks cited in the announcement.
SourcesGPT-6.1 TerraOpenAI · 2026-10-26Read the sourceReported gains on professional task evaluations
OpenAI reports that at the medium reasoning setting, GPT-6.1 Terra scored 15 percentage points higher than GPT-5.6 Terra on Agents’ Last Exam, an evaluation that includes tasks such as turning architectural plans into 3D models. These results are reported by the model provider.
SourcesGPT-6.1 TerraOpenAI · 2026-10-26Read the sourceSystem card specifies evaluation conditions
According to the system card, evaluations took place in research environments or through the API, where system prompts, available tools and reasoning settings may differ from those in production ChatGPT. OpenAI treats the model as Critical for cybersecurity and High for biological and chemical capabilities, applying the same class of safeguards used for Sol and Astra.
SourcesAddendum to GPT-6 Astra System Card: GPT-6.1 TerraOpenAI Deployment Safety Hub · 2026-10-26Read the sourceH/ACC perspective
Shortening the loop between seeing and acting
H/ACC editorial interpretation: connecting visual understanding to computer use can shorten the loop between observing an interface, taking action and checking the result. If improvements in speed and cost carry over to real workflows, repeated checks and timely corrections can more easily become routine.
H/ACC editorial interpretation: practical results also depend on how tools, permissions and completion criteria work together. Recording actions, checking outcomes and retaining explicit authorization for steps involving real assets are necessary to turn the model’s ability to act into a traceable workflow.
Limits of the evidence
- The provider’s benchmark scores are not real-world business success rates. Speed, cost and completion quality may vary with the task, tools and reasoning setting.
- The system card still reports failures in some boundary and authorization evaluations. Capability ratings and safeguard descriptions are not guarantees of zero risk.
- H/ACC has not independently reproduced the model evaluations. This article does not translate test-environment performance or per-token pricing directly into gains for every workflow.
What to watch nextAI-assisted analysis
- With the same tools, tasks and reasoning setting, how do completion rates, rework cycles and total cost per task change in real workflows?
- When facing malicious instructions on web pages or steps requiring authorization, such as login and payment, can the system stay within operational boundaries and leave a traceable record?
Sources
Reporting and research used in this story.
- GPT-6.1 TerraOpenAI · Source published 2026-10-26
- Addendum to GPT-6 Astra System Card: GPT-6.1 TerraOpenAI Deployment Safety Hub · Source published 2026-10-26