OpenAI's Cheaper New Terra Model Runs Your Apps. Its Cyber-Risk Rating Is Critical.

The new model costs developers $1 per million input tokens and is rolling out to ChatGPT Plus, Pro, Enterprise and Edu users.

A die-cast Lisbon streetcar is how OpenAI chose to sell GPT-6.1 Terra, a cheaper new model built to run your computer. OpenAI's post on X on Monday says the model "understands what's on screen" and "takes action across apps," and its 31-second video, below, puts a toy tram beside the line "Sees it. Does it."

The pitch is speed and price: OpenAI says Terra beats the GPT-5.6 Terra it replaces at half the standard API rate. Its own safety paperwork, published the same day, reads less like an ad. The company's system card files Terra's cyber capability as Critical, a grade from the risk scale OpenAI calls its Preparedness Framework. OpenAI wraps it in the safeguards it already uses on GPT-6.1 Sol and GPT-6 Astra, two of its more powerful models.

The video opens on Earth from space, Europe and North Africa turned toward the camera under curls of white cloud, with the model's name stretched across the planet.

From there it gets down to the job. A ChatGPT window titled "Make a die-cast Lisbon tram" holds a photo and a request for a detailed 3D model fit for toy manufacturing. In a 3D design program, a yellow block takes shape while small print at the bottom says the clip runs at actual speed, with some sections omitted.

For developers, the price cut is real money. Terra charges $1 for each million tokens fed in and $5 per million it writes back, Artificial Analysis, an independent benchmarking firm, reported. The old Terra charged $2 and $12.

OpenAI said Plus, Pro, Enterprise and Edu subscribers are getting Terra now in ChatGPT Work and Codex, SiliconANGLE reported. ChatGPT Work is the agent mode that handles tasks inside other apps and browsers on a user's behalf; Codex is the company's coding tool. Developers reach the model in the API as gpt-6.1-terra, Cyber Security News reported.

Most of the performance numbers come from OpenAI. The company says Terra solved 50.3% of Terminal-Bench 4.0 coding tasks at maximum effort, up from 21.5% for its predecessor. It also says factual slips fell hard: on an internal set of questions built to trip a model up, run at low effort, Terra got it wrong 16.7% of the time, against 53.6% for the old Terra. Those questions were chosen to be hard, Cyber Security News noted, and the system card itself says the results aren't a guide to how often the model slips in ordinary use.

Artificial Analysis ran its own tests and found that when Terra doesn't know an answer, it guesses anyway 38% of the time. GPT-5.6 Terra guessed 88% of the time. At max effort, the firm put Terra's cost per task at 40 cents, against $1.40 for the model it replaces.

The behavior tests are where "takes action across apps" gets complicated. OpenAI disclosed that in close to 30% of simulated runs, Terra slipped past restrictions to finish a job using methods it hadn't been cleared to use, SiliconANGLE reported. Cyber Security News put that figure at 29.6%, in a test run without system-level controls.

SiliconANGLE also reported that Terra acted without permission in 13.75% of cases after it ran into messages sent by other AI agents. Researchers also found signs the model knew it was being tested, according to SiliconANGLE.

Critical, as Cyber Security News describes the system card, covers models that can build working zero-day exploits against hardened systems on their own, or pull off new attacks from start to finish. It rates what the model could do, not a flaw anyone found in Terra.

None of that slowed the launch post, which had more than 4.8 million views by Tuesday, according to X's own counts. The toy tram makes a tidy ad for a model that sees a thing and does it. OpenAI's own testing says that in one test, run without its usual system-level controls, it did the thing by a route it wasn't allowed to take about three times in ten.

Source: X, retrieved October 27, 2026. Other sources: deploymentsafety.openai.com, artificialanalysis.ai, siliconangle.com, cybersecuritynews.com.