GPT-6.1 Terra: Operates Your Computer, Costs a Tenth of the Flagship per Token

OpenAI has released a new model, GPT-6.1 Terra. It's officially positioned as a fast, efficient mid-tier offering, packing visual reasoning, computer control, and writing into a single model. The goal is to handle everyday office scenarios like drafting proposals, reading technical drawings, filling out forms, and editing videos.

Terra is now available to ChatGPT Plus, Pro, Enterprise, and Edu users, and has also been integrated into ChatGPT Work and Codex. Developers can access it via API under the model name gpt-6.1-terra.

Its most striking feature is the ability to operate a computer on its own. On the desktop control benchmark OSWorld 2.1, it beats the best score of GPT-5.6 Terra at medium reasoning effort. The previous generation had to push to its limits; this one achieves it in its default state.

The pricing makes it even more interesting. On the Agents' Last Exam benchmark, Terra at medium effort scores 15% higher than GPT-5.6 Terra, with a per-task cost only one-third of the older model, while matching the flagship GPT-6 Astra on score—yet its per-token price is just one-tenth of the flagship. At maximum effort, its token efficiency is over three times that of Claude Opus 5.5, at roughly a quarter of the price, translating to a per-task cost about 9 times lower. It also completes tasks more than three times faster than both GPT-6 Astra and GPT-6.1 Sol.

Specific data:

Benchmark	Result
OSWorld 2.1 (Desktop Control)	Exceeds GPT-5.6 Terra's best score at medium reasoning effort
Agents' Last Exam (Agent Capabilities)	Scores 15% higher at medium effort, with per-task cost around 1/3 of the older model
MMMU-Pro (Visual Reasoning)	Surpasses GPT-5.6 Terra across all effort levels; at low cost, approaches the two flagship models
Factuality Test	Error rate drops from 53.6% to 16.7% at low effort, and to 4.3% at high effort
Terminal-Bench 4.0 (Programming)	Solves 50.3% of tasks at high effort, more than double the older model's public result

Capabilities have grown, and so have the guardrails. Terra has achieved a Critical rating in OpenAI's internal cybersecurity evaluation, inherits GPT-6 Astra's protections, is better at staying within task boundaries, proactively discloses its limitations, and is more resistant to prompt injection.

One signal worth noting: the gap between flagship and mid-tier is narrowing. Previously, cheaper usually meant notably weaker. Now the mid-tier delivers comparable results at a fraction of the flagship's cost, and faster too. For people building Agent applications, the cost structure may need a rethink. OpenAI is using price competition to capture the Agent entry point.