GPT-6.1 Terra

A Different Kind of Mid-Tier Model

GPT-6.1 Terra is not positioned simply as a smaller GPT-6.1 Sol.

Its product identity is more specific.

OpenAI introduced the model on October 26, 2026 around the idea of moving from visual understanding to action.

The model is designed for work such as:

• matching product photos to catalog entries;
• interpreting technical drawings;
• understanding application screens;
• filling forms;
• entering invoice data into accounting software;
• editing video through computer interfaces;
• drafting and revising natural-language content;
• running coding and terminal workflows.

This gives Terra a distinctive place in the GPT-6 family.

GPT-6 Luna targets the lowest-cost focused high-volume work.

GPT-6.1 Terra targets visually grounded computer work and communication at scale.

GPT-6.1 Sol targets more complex professional, coding, and computer-use projects.

GPT-6 Astra remains the maximum-capability tier.

Terra combines:

• 1,050,000-token context;
• 128,000 maximum output tokens;
• text and image input;
• text output;
• reasoning from none through max;
• function calling;
• structured outputs;
• computer use;
• search and file tools;
• coding tools;
• prompt caching.

Its knowledge cutoff is June 1, 2026.

Computer-Use Quality Is More Than Click Accuracy

A reliable computer-use agent needs to inspect the current interface state, understand the user’s objective, identify relevant controls, choose the correct action, observe the result, recover from unexpected state, and continue until the task is complete.

OpenAI evaluates Terra on Agents’ Last Exam, a benchmark built around professional computer-based work.

Example tasks include:

• converting architectural drawings into a 3D model;
• reviewing an animated film shot by shot;
• creating a time-coded restoration log.

At medium reasoning effort, OpenAI reports that Terra scores 15 percentage points higher than GPT-5.6 Terra at about one-third the cost.

OpenAI also says Terra performs on par with GPT-6 Astra on the benchmark while costing only one-tenth as much per token.

Speed is part of the positioning.

On these tasks, Terra completes work more than three times faster than both GPT-6 Astra and GPT-6.1 Sol.

That combination—visual understanding, UI action, speed, and lower inference cost—is the main reason to evaluate Terra separately from Sol.

A workday automation rarely ends with a click.

The same workflow may need to draft a proposal, rewrite a customer-facing message, summarize completed work, explain an exception, prepare notes, update documentation, or create follow-up communication.

Terra is designed to combine these writing tasks with the visual and computer-action parts of the workflow.

A single model can inspect an interface, perform the action, draft the result, and revise the wording.

Writing quality should still be evaluated on real product requirements.

Measure instruction adherence, tone, factual consistency, unnecessary verbosity, revision quality, and preservation of supplied facts.

50.3% on Terminal-Bench 4.0

Terminal-Bench 4.0 evaluates coding agents operating through the command line.

Tasks include practical engineering problems such as fixing a webpage whose layout shifts while loading.

At maximum reasoning effort, OpenAI reports:

• GPT-6.1 Terra: 50.3%;
• GPT-5.6 Terra: 21.5%.

Terra therefore more than doubles the published GPT-5.6 Terra result.

This makes the model relevant for:

• bounded bug fixes;
• terminal-based repair;
• frontend debugging;
• code transformation;
• test-driven iteration;
• infrastructure tasks;
• coding assistance that includes tools.

However, OpenAI explicitly says that GPT-6.1 Sol and GPT-6 Astra remain the better choice for the most complex coding tasks.

This suggests a practical routing policy.

Use Terra for frequent coding tasks, visually grounded debugging, bounded repository work, and workflows where latency and cost matter.

Escalate when architecture is highly ambiguous, the repository change is broad, repeated repair attempts fail, the task requires deeper research, or the cost of a wrong implementation is high.

Cost Is Central to Terra’s Positioning

Standard rates are:

• input: $1.00 per 1M tokens;
• cached input: $0.05 per 1M tokens;
• 30-minute cache write: $1.25 per 1M tokens;
• 12-hour cache write: $2.50 per 1M tokens;
• output: $5.00 per 1M tokens.

Cached input costs 5% of normal input.

OpenAI positions this pricing as part of Terra’s value proposition.

The model is 50% cheaper than GPT-6.1 Sol on standard input and output token rates.

Processing tiers change the economics further.

OpenAI lists:

• Batch: 50% below Standard;
• Flex: 50% below Standard;
• Fast: 2× Standard.

Regional processing adds a 10% premium where available.

Tool-specific calls can have separate charges.

medium is the default.

This differentiates Terra from GPT-6.1 Sol, which does not support a true none reasoning mode.

That matters for high-volume workflows.

A form-filling or extraction task may not need the same reasoning budget as a technical drawing analysis, difficult computer-use trajectory, coding repair, or scientific workflow.

A useful evaluation strategy is:

1. start with medium;
2. test lower effort for routine work;
3. test higher effort for difficult cases;
4. compare success rate, latency, tokens, and cost.

OpenAI’s migration guidance specifically recommends comparing GPT-6.1 Terra and GPT-5.6 Terra at the same reasoning effort on representative images and workflows.

What They Share

Both models provide:

• 1.05M context;
• 128K maximum output;
• text and image input;
• Responses API tools;
• computer use;
• function calling;
• structured outputs.

Where Terra Is Different

GPT-6.1 Terra:

• $1 input;
• $5 output;
• supports none reasoning;
• optimized for visual reasoning;
• optimized for precise computer use;
• designed for high-volume workday automation;
• substantially faster on OpenAI’s Agents’ Last Exam evaluation.

Where Sol Is Different

GPT-6.1 Sol:

• $2 input;
• $10 output;
• starts at low reasoning;
• positioned for near-Astra complex work;
• better fit for the hardest coding and deep professional tasks.

OpenAI explicitly says Sol and Astra remain the better choices for the most complex coding workloads.

The correct architecture may use both.

Route:

• routine visual/computer work → Terra;
• difficult coding or professional reasoning → Sol;
• maximum-capability cases → Astra.

OpenAI says GPT-6.1 Terra is more likely than GPT-5.6 Terra to:

• stay within the user’s requested scope;
• disclose limitations;
• disclose blockers;
• resist malicious prompt-injection attempts.

This is especially important for computer-use agents.

A model operating software may encounter untrusted content inside webpages, documents, emails, application fields, and retrieved files.

Prompt-injection resistance reduces risk but does not replace application controls.

Use:

• scoped permissions;
• tool allowlists;
• human confirmation for consequential actions;
• sandboxing;
• logging;
• output validation.

OpenAI also states that GPT-6.1 Terra reaches the Critical threshold in cybersecurity under its Preparedness Framework and uses the safeguard stack developed for GPT-6 Astra.