More capable AI.
A lower cost to act.

The development that matters

GPT-6.1 Terra succeeds GPT-5.6 Terra with improvements in image understanding, desktop interaction, writing and coding. It targets high-volume work such as reading invoices, interpreting technical drawings and transferring information between applications.

For businesses, the opportunity is a lower-cost model for routine agent steps, with stronger models reserved for difficult planning and debugging. Its release does not by itself make an end-to-end workflow dependable.

Read the official announcement ↗

Where it could be useful

• Document processing: extract structured information from invoices, scanned forms and visual documents.
• Application workflows: navigate interfaces, prepare form entries and inspect on-screen results.
• Professional writing: draft proposals, reports and customer-facing material.
• Coding: handle scoped terminal tasks and implementation work, escalating harder problems when necessary.

Computer use requires an agent environment and connected tools. The model alone cannot operate a user's computer or access business systems.

What the evidence says

OpenAI reports a 50.3% score on Terminal-Bench 4.0 at maximum reasoning effort, compared with its published 21.5% for GPT-5.6 Terra. The company also reports substantial gains on professional computer-use tasks and lower factual error rates in internal evaluations.

These are vendor-reported results, not independent proof of performance in production. Benchmark outcomes depend on reasoning settings, tools and the surrounding agent setup. OpenAI still positions Sol and Astra as stronger choices for the most complex coding tasks.

Limits that still matter

The system card reports improvements over older models, alongside higher flag rates than Sol or Astra in some categories, including destructive actions and inbound prompt injection. Strong aggregate results should not be read as a guarantee of safe behavior.

Production deployments still need constrained permissions, validated tool arguments, idempotent writes and checks that actions actually succeeded. Browser automation should complement reliable APIs where direct integration is available.

Terra is a hosted, proprietary model. No downloadable weights or local Ollama version were announced; local GPU capacity does not determine access.

What to test first

• Run a document extraction comparison. Use the same anonymized samples and schema across models; measure field accuracy, latency and cost per correct result.
• Start computer-use trials with read-only tasks. Test navigation, error recognition and recovery before enabling consequential actions.
• Measure completed workflows, not token prices alone. Retries, reasoning tokens and tool charges can change the economics.