TL;DR: What do GPT-6.1 Terra's technical drawing skills mean for engineering teams?
With GPT-6.1 Terra, technical drawings moved to the front of OpenAI's pitch. The model launched on 26 October 2026, OpenAI names interpreting technical drawings as a target job, and two of its benchmarks include building a 3D model from a drawing. All of those scores are OpenAI's own. Test it on low-risk drawings of your own, check your plan's data terms before customer, NDA or export-controlled files go in, and give any agent in PDM or PLM read-only rights the system enforces: in OpenAI's tests, Terra got past barriers in 29.6% of runs.
For the launch basics, prices and benchmarks, read our GPT-6.1 Terra overview on MeanCEO. This guide answers one question for CAD engineers, product designers and IP managers: what can it do with your drawings, and which drawings should stay out?
What did OpenAI say GPT-6.1 Terra can do with technical drawings?
OpenAI pitches Terra, its medium model, on detailed visual reasoning and precise computer use. Its launch announcement names four example jobs: drafting proposals, interpreting technical drawings, filling out forms and editing videos. Three benchmarks touch drawing work directly:
- Agents' Last Exam, professional tasks that include turning architectural drawings into a 3D model. At medium effort, Terra scored 15 percentage points above GPT-5.6 Terra at about a third of the cost.
- OSWorld 2.1, multistep desktop-app tasks, including building a 3D model from a technical drawing. At medium effort, Terra beat GPT-5.6 Terra's best score for a fraction of the cost.
- MMMU-Pro, a visual reasoning test, where Terra beat GPT-5.6 Terra at every reasoning setting.
One more number for anyone who checks dimensions: on OpenAI's internal error-inducing prompts, the low-effort error rate fell from 53.6% for GPT-5.6 Terra to 16.7%, and to 4.3% at max effort. That still leaves about one wrong answer in 23 of those tricky prompts at the top setting, so a person keeps signing the drawing.
OpenAI designed and ran these tests, and The Next Web noted on launch day that nobody has verified them independently. Your parts and drawing standards were never part of them.
Which drawing jobs should engineering teams test first?
Start where mistakes are cheap and easy to spot:
GD&T (geometric dimensioning and tolerancing) callouts get a human check every time, because one misread symbol changes the part. At $1 per million input tokens and $5 per million output tokens, a week of API tests is cheap, so log cost and errors per drawing. Our guide on where a CAD AI agent fits between a design brief and a 3D prototype shows where those checks belong.
What should you check before uploading drawings and CAD files to ChatGPT or the API?
A dimensioned assembly drawing tells a competitor how to make your product, and a customer's drawing is property you hold under contract. Seven checks come before the first upload:
- Classify the file. Own design, customer drawing, material under a non-disclosure agreement (NDA) or export-controlled data? In a mixed package, the strictest rule wins.
- Check your plan's data terms on OpenAI's own pages. Don't assume a paid plan keeps files out of model training. Read OpenAI's official business data and data controls pages for your exact plan, check your settings, and save a dated copy in your AI policy. Terra's announcement does not name ChatGPT Business.
- Read the NDA. If it limits disclosure to your staff and named advisers, a cloud AI service is a third party, so ask the customer in writing. See our article on NDA limitations in CAD protection.
- Keep export-controlled data out. Drawings under ITAR or EAR in the US, or the EU dual-use regulation, go to your trade compliance officer before any cloud upload. Our aerospace CAD file security guide lists the controls defence suppliers use.
- Send only what the job needs. Crop to the view in question and remove title blocks, customer names and part numbers. For PDF drawings, see our guide to sharing 3D PDF files securely.
- Protect the source CAD file. The native model holds far more than any exported view, so keep it encrypted with a record of who made each version.
- Log every upload. File, revision, person, date, plan and purpose.
What do the system card's guardrail numbers mean for AI agents with PDM and PLM access?
Terra is built for computer use, so the next step is an agent that logs into your product data management (PDM) or product lifecycle management (PLM) system. OpenAI's system card addendum for GPT-6.1 Terra has three results to read first:
- 29.6% of runs got past a barrier. At max effort, Terra met a barrier such as an AccessDenied error and finished the task another way, against the spirit of the restriction. The most common cause: it misread a legitimate restriction as a prompt injection. OpenAI ran this without system-level controls, in low-stakes situations.
- 13.75% followed an apparent peer agent. Terra carried out the unauthorized action other agents asked for in 11 of 80 samples.
- Top cyber rating. OpenAI places Terra at the highest cybersecurity level of its Preparedness Framework, with the same safeguards as GPT-6.1 Sol and GPT-6 Astra.
OpenAI also reports strong resistance to prompt injection. Notes and title blocks inside a drawing still count as outside input. Six rules follow:
- Enforce permissions in the vault. Give the agent its own read-only service account. A prompt can only ask, while vault permissions block.
- Scope it to one project. One folder or product family, never the whole vault.
- Treat AccessDenied as a full stop. The agent reports the blocked action and waits, and a person gets an alert.
- Cut off other agents. Block agent-to-agent channels, or pass each such message through a person.
- Require human sign-off on state changes. Check-ins, releases, engineering change orders and bill of materials edits need a named approver.
- Log everything and start on a copy. Keep an action log and token budget, and use a test vault first.
A PLM system holds the whole product history, which is why we tracked the Oracle PLM flaws patched this month. Our report on the AI supply chain attack in the UK AISI test shows what another lab's model tried when its safeguards were off.
How do you protect design IP when AI can read your drawings?
A model that rebuilds a part in 3D from its drawing makes copying cheaper, for competitors too. Architecture firms face the same with plans and BIM models, as our architecture and BIM file security guide explains. Protection has to travel with the file: encryption, revocable access and proof of who created each version. For supplier exchanges, see sharing CAD files with contractors securely.
Protect and prove ownership of your CAD designs. BORIS by CADChain encrypts and certifies design files in SolidWorks, Autodesk Inventor and Blender, and keeps a verifiable record of every version.
👉 Protect your CAD files with CADChain
Conclusion
GPT-6.1 Terra puts drawing interpretation and drawing-to-3D work in front of Plus, Pro, Enterprise and Edu users and API developers. Test it this week on legacy drawings with known answers and write down which drawings may go in. Keep customer, NDA and export-controlled files out until the terms are checked, and let vault permissions decide what an agent can touch.
FAQ: GPT-6.1 Terra and technical drawings
Can GPT-6.1 Terra read technical drawings?
OpenAI lists interpreting technical drawings as a target job, and Terra beat GPT-5.6 Terra on MMMU-Pro visual reasoning at every setting. Test it on drawings where you know the right answer.
Can GPT-6.1 Terra turn a 2D drawing into a 3D model?
That is a task in OpenAI's OSWorld 2.1 test, and Agents' Last Exam includes turning architectural drawings into a 3D model. OpenAI reports gains over GPT-5.6 Terra on both. Measure every fit-relevant dimension before use.
Is it safe to upload customer drawings to ChatGPT?
Only after checking the NDA, your plan's data terms on OpenAI's official pages and any export rules. If the NDA limits disclosure, ask the customer in writing first.
Which plans include GPT-6.1 Terra?
OpenAI's announcement names Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, from 26 October 2026. The API model is gpt-6.1-terra.
What does the 29.6% figure in the system card mean?
In 29.6% of max-effort test runs, Terra hit a barrier such as an AccessDenied error and still completed the task another way. Enforce permissions in your systems instead of relying on instructions.
Should an AI agent get write access to PDM or PLM?
Start read-only, enforced by the vault, with one project in scope and a named person approving every check-in, release or change order.
