OpenAI Launches GPT-6.1 Terra: Half-Price Model for Computer Use
It cuts API pricing in half and runs UI actions three times faster than flagship models, but lower general coding and reasoning marks mean it will not replace Sol or Astra across the board.
On Oct. 26, 2026, OpenAI introduced GPT-6.1 Terra, a mid-tier model tuned specifically for day-to-day workflows and computer use. Positioned between lightweight Luna and frontier-class Astra, Terra tackles routine agent automation. OpenAI halved token(a small piece of text, roughly part of a word, that AI models read and write) pricing compared to both previous-generation GPT-5.6 Terra and upper-tier GPT-6.1 Sol while sharpening vision and screen navigation capabilities.
The model is available via the API (gpt-6.1-terra), inside ChatGPT paid tiers (Work, Plus, and Pro), and across partner ecosystems including GitHub Copilot and Amazon Bedrock.
Targeted at Visual Reasoning and Screen Actions
Terra prioritizes blueprint parsing, screen analysis, and hands-on desktop interaction. OpenAI separated these workloads during training to bypass the latency and high costs of running top-tier frontier models on routine computer tasks.
The model reads dimensions from architectural drawings, extracts key fields from long screenshots, and navigates video frames. In computer use workflows—filling out web forms, parsing invoices, and driving desktop applications via mouse and keyboard—Terra completes tasks three times faster than Astra or Sol.
On the Agents’ Last Exam benchmark, Terra scored roughly 15 percentage points higher than GPT-5.6 Terra at medium reasoning effort, closing the gap with OpenAI’s flagship Astra.
Halved API Pricing and 1.1M Token Context
API rates sit at roughly half the cost of neighboring models. The context window supports up to 1.1 million (1,050,000) input tokens and 128,000 output tokens.
| Model | Input / 1M tokens | Cached input / 1M tokens | Output / 1M tokens | Context window |
|---|---|---|---|---|
| GPT-6.1 Terra | $1.00 | $0.05 | $5.00 | 1.1M tokens |
| GPT-5.6 Terra | $2.00 | $0.10 | $10.00 | 256K tokens |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | 1M tokens |
| GPT-6 Astra | $10.00 | $0.50 | $50.00 | 1M tokens |
Developers can adjust reasoning_effort from none to max. Setting it to none unlocks classic sampling parameters like temperature and top_p, letting teams trade exploratory reasoning for minimal latency.
Where It Fits and Where It Falls Short
OpenAI researcher Ted Sanders clarified Terra’s scope on developer forums.
Terra matches or edges out top-tier models on visual parsing, UI automation, and technical writing. However, it still lags behind Sol and Astra on algorithmic proofs, high-difficulty mathematical reasoning, and deep architectural refactoring.
Workflows that demand multimodal inputs and rapid iterative calls—such as invoice pipelines and browser automation—fit Terra well. In contrast, complex systems engineering and frontier research remain better suited for Sol or Astra despite higher token bills.
Clarifications and Easy Misunderstandings
- Not a standalone app: “ChatGPT Terra” is not a separate application. GPT-6.1 Terra operates as a model option within the API and existing paid ChatGPT tiers.
- Not a blanket benchmark winner: While fast at screen tasks and technical communication, Terra trails Sol and Astra on broad logic and coding benchmarks.
- Prompt caching has conditions: The $0.05 per million token cache rate applies only when consecutive requests share an identical prompt prefix.
What to Watch Next
The computer use segment is now a direct race between Anthropic’s Claude Computer Use and OpenAI’s Terra line. As mid-range pricing drops to $1 per million input tokens, whether cost-efficient models can reliably anchor enterprise automation pipelines without excessive oversight remains the key test.
Sources
- Introducing GPT-6.1 Terra (OpenAI)
- Hacker News Discussion on GPT-6.1 Terra (Hacker News)
Source: openai.com