What happened
Microsoft announced on October 26, 2026 that OpenAI’s GPT-6.1 Terra is generally available in Microsoft Foundry. Microsoft describes Terra as a quick, efficient model for day-to-day professional work, particularly tasks that begin with visual analysis and end with an action in software. Its stated strengths are visual reasoning, faster and more precise computer use, and clear writing.
The model accepts text and images and supports up to 1 million input tokens. Microsoft says improved visual reasoning is intended to help it identify relevant details in product photos, technical drawings, floor plans, and application screens. For computer-use work, the company says Terra can pair what it sees with actions across browsers, desktop software, and apps. It also positions the model for plain-language summaries and drafts of emails and proposals.
For developers, Terra exposes controls for reasoning effort, including a none setting for latency-sensitive calls, and for output verbosity. It can also send interim messages while it is still working when the orchestration layer supports that behavior. Microsoft says Terra works through the existing Foundry Responses API, with no new SDK requirements or schema changes.
Availability starts in Foundry Models through Global Standard deployments in all supported regions. Data Zone Standard deployments are available in the US, EU, and APAC. Microsoft frames this as a way to use the model where an application is already built and scaled while retaining data-residency choices. Its examples include completing forms, reviewing campaign assets, reading technical drawings, updating retail listings from supplier images, and finding and editing a moment in video software.
Microsoft positions Terra alongside GPT-6 Astra rather than as its replacement: Astra is for open-ended, multi-step work and planning, while Terra is for high-volume visual and screen-based tasks at a lower price point. In Global Standard short-context pricing, Terra costs $1.00 per million input tokens, $0.05 cached input, $1.25 cached writes, and $5.00 output. Long-context Global Standard pricing is $2.00 input, $0.10 cached input, $2.50 cached writes, and $7.50 output per million tokens. Data Zone pricing is higher and varies by region.
Microsoft also points to its Decisions API for steps that require a fast structured answer, such as classification or routing. The company’s proposed division is Terra for visual and on-screen work, with the Decisions API making the next routing decision.
What the archive adds
This is a new availability event in a known rollout, not a restatement. Terra was already generally available in GitHub Copilot across IDEs, CLI, and cloud agent and live through Vercel AI Gateway. Foundry adds Microsoft’s model platform, its deployment geography, and its governance-oriented production environment.
Trend signal
This supports Harness architecture outlives model choice. Microsoft emphasizes that Terra can enter existing Foundry APIs and tools without an SDK or schema change, while exposing deployment, evaluation, and governance capabilities around it. That does not prove model interchangeability, but it shows the platform layer being presented as the durable place to operationalize a newly available model.
Connections to other events
The Foundry release follows Terra’s availability in GitHub Copilot across IDEs, CLI, and cloud agent and through Vercel AI Gateway. Those are different operating surfaces: Copilot is a developer-workflow product, AI Gateway is a gateway route, and Foundry is Microsoft’s platform for building and scaling applications.
Why it matters
For a team already using Microsoft Foundry, the useful change is not merely another model name in a catalog. Terra is available through the Responses API and the surrounding platform controls teams may already use, so adding it does not require a separate integration shape. The release also makes model selection more specific: visual inspection and UI action can be assigned to Terra, while open-ended planning can remain with Astra.
The published pricing and deployment variants make this an implementation decision rather than a vague capability claim. Teams can estimate the cost of image-and-screen workflows, decide whether Global Standard or a Data Zone deployment fits their data-residency needs, and test whether the latency trade-off of lower reasoning effort is acceptable for a bounded step. The model’s availability does not remove the need to validate screen actions, but it gives Foundry builders a concrete option for that class of task.
What happens next
- Teams already running Foundry agents will test Terra on visual inspection and screen-action steps. (3 months, 70%). Watch for: Foundry examples or customer deployments that report measured task outcomes.
- More implementations will split planning from visual execution across model roles. (6 months, 55%). Watch for: reference architectures pairing Terra with a planning model and structured routing.
Limits and uncertainty
Microsoft’s post is a product announcement, not an independent evaluation. It claims gains in visual reasoning and computer-use speed but provides no benchmark results, task-success rates, latency measurements, or comparison against specific alternatives. The listed workflow examples are illustrations rather than evidence that an agent can safely complete them without review.
The announcement also does not state the practical limits of computer use in particular applications, browsers, or desktop environments. A 1 million-token input limit does not establish effective performance over that full context, and the supplied pricing table does not determine the total cost of an agent workflow, which also depends on images, retries, tool calls, output volume, and orchestration.
Skill to build now
Build a small evaluation set for one visual-to-action step in your workflow. Use real but safe examples such as catalog-image matching or extracting an issue from a UI screenshot, define the expected action and a human-verifiable completion check, then run the same set with different reasoning-effort settings. Record task completion, correction time, latency, and token use. This turns Microsoft’s positioning into a deployment choice grounded in your own work rather than a general model claim.
Reader takeaway
GPT-6.1 Terra is now a production option inside Microsoft Foundry for agents that must interpret images or screens and then act in software. Its main immediate value is integration: existing Responses API users can test a specialized visual-and-computer-use model with regional deployment choices and explicit token pricing.
Open question
Can Terra’s claimed gains in visual reasoning and computer use translate into reliably verified completion on real enterprise interfaces, rather than only faster generation or navigation?