GPT-6.1 Terra explained: visual reasoning, computer use, writing and the price cut from GPT-5.6 Terra

OpenAI released GPT-6.1 Terra in its API changelog on October 26, 2026, with the model identifier gpt-6.1-terra. It is positioned as a fast, efficient model for visual reasoning, precise computer use and clear, natural writing.

For developers, pricing matters alongside the intended workloads. Current Standard text rates for short-context requests are $1 per million input tokens and $5 per million output tokens, below the $2 and $12 listed for GPT-5.6 Terra. Lower token prices do not automatically produce the same percentage reduction in cost per completed task.

Information was checked on October 27, 2026 against official OpenAI documentation. TboxIO has not independently benchmarked the model or measured its speed. Cost examples are calculations from published rates, not actual bills.

What is GPT-6.1 Terra designed for?

Official examples include matching product photos to catalog entries, interpreting or comparing technical drawings, acting on application screens, and drafting or revising text. These tasks require the model to understand visual or interface details and turn that understanding into a decision, action or written result.

An ecommerce team could evaluate matching accuracy among similar product photos. A software team could assess recognition of controls, status messages and next actions. An editorial team could check whether revisions preserve meaning and sound natural. These are evaluation ideas based on the stated uses, not proven outcomes for every industry.

OpenAI positions GPT-6 Astra for the hardest reasoning, coding and professional work; GPT-6.1 Sol balances strong capability and cost; Terra emphasizes efficient visual reasoning, computer use and writing; GPT-6 Luna targets focused, high-volume tasks. Astra or Sol benchmark results must not be attributed to Terra simply because they share the GPT-6 family.

Key specifications: 1.05M context, text and image input

Item	Official GPT-6.1 Terra information
API identifier	gpt-6.1-terra
API changelog release date	October 26, 2026
Context window	1,050,000 tokens
Maximum output	128,000 tokens
Knowledge cutoff	June 1, 2026
Input	Text and images
Output	Text
Audio and video	Listed as unsupported on the model page
Reasoning effort	none, low, medium, high, xhigh, max; medium by default
Streaming and structured outputs	Supported
Fine-tuning	Listed as unsupported
Tool calling	Requires the Responses API

The context window and maximum output are separate limits. They cannot be added together, and the output limit does not imply that every response will be that long. A knowledge cutoff is also distinct from access to current information: up-to-date claims still require appropriate retrieval and source checks.

Accepting images and producing text does not mean that the model itself outputs images. The official tool list includes image generation as a tool capability; distinguish native model modalities from tools used by a workflow.

What changes from GPT-5.6 Terra can be confirmed?

Item	GPT-5.6 Terra	GPT-6.1 Terra
Official positioning	Balance of intelligence and cost	Efficient visual reasoning, precise computer use and natural writing
Context window	1,050,000	1,050,000
Maximum output tokens	128,000	128,000
Knowledge cutoff	February 16, 2026	June 1, 2026
Text/image input and text output	Supported	Supported
Reasoning effort	none through max; medium default	none through max; medium default
Standard text input / 1M tokens	$2.00	$1.00
Standard cached input / 1M tokens	$0.20	$0.05
Standard text output / 1M tokens	$12.00	$5.00

The comparison uses the Standard text rates listed on the two model pages when checked, for requests with no more than 272K input tokens. Input rates fall 50%, cached-input rates 75%, and output rates approximately 58.3%. These are rate changes, not quality improvements or guaranteed savings for every request.

Context and output limits are unchanged, so this is not a context-window expansion. Computer use, image input and tool support should not be described as newly introduced merely because the new page lists them. The useful comparison is quality, action accuracy, latency and total cost on the same tasks.

Benchmarks and speed: what can we conclude?

The official changelog, Terra model page and guide checked for this article do not provide a directly citable table of Terra-specific standardized benchmark results or a fixed tokens-per-second figure. We therefore do not invent SWE, vision or computer-use scores, or claim a multiple of the previous model's speed.

Product positioning identifies what to test; it does not replace results. Relative capability indicators on a model page cannot be converted into OSWorld or SWE-bench scores.

Use a fixed set of representative tasks. For product matching, record correct and incorrect matches. For drawings, check key dimensions and labels. For computer use, record task completion, wrong clicks and human takeovers. For writing, evaluate factual preservation, constraints, naturalness and editing effort. Also record end-to-end time, input/output tokens, tool calls and retries.

API pricing: long context and cache writes matter

The following prices are in USD per million text tokens, using Standard processing and no more than 272K input tokens.

Billing item	GPT-6.1 Terra
Uncached input	$1.00
Cached input read	$0.05
30-minute cache write	$1.25
12-hour cache write	$2.50
Output	$5.00

Cheap cache reads do not mean free cache creation. Thirty minutes and twelve hours are different retention choices with different write rates. Savings depend on stable shared prefixes, subsequent reuse and actual cache hits.

• Above 272K input tokens, input and cache rates double and output rates increase 1.5 times for the entire request, not just the portion beyond the threshold.
• The model page lists Batch and Flex at 50% below Standard and Fast at twice Standard. Check each tier's timing, speed and eligibility separately.
• Regional processing carries a 10% premium where applicable; calculate using the actual endpoint and conditions.
• Search, containers and file search can add fees. Multi-turn agents also accumulate repeated input, tool results and retries, so a single Q&A estimate is not the cost of a full agent task.

For a purely textual example, suppose a request uses 10,000 uncached input tokens and 2,000 billable output tokens, no tools or cache, Standard processing and short-context rates. The model-token cost is approximately $0.02 with GPT-6.1 Terra, versus $0.044 at the listed GPT-5.6 Terra rates: about 54.5% lower for this particular token mix.

This fixes token counts for illustration. Real reasoning tasks may bill more output than the visible final text; inspect reasoning and output usage. Image input should not be assumed to have one fixed price per image or inserted unchanged into this text-only example.

Integration: move tool workflows to the Responses API

The official guide recommends comparing GPT-5.6 Terra and GPT-6.1 Terra on representative images and workflows at the same reasoning effort. If an application currently calls tools through Chat Completions, migrate those requests to the Responses API. Seeing both endpoints in release information does not imply equivalent tool-calling support through Chat Completions.

Terra supports none, low, medium, high, xhigh and max, defaulting to medium. Start there, increasing effort only when quality gains justify the additional time and tokens. Lower effort can be evaluated for latency-sensitive workloads, but neither none nor max is automatically best for every task.

The Responses tool list includes web search, file search, code interpreter, hosted shell, apply patch, Skills, computer use, MCP and tool search. Model support does not mean integration is already complete: permissions, execution environments, error handling, state management and verification depend on the application.

Computer use particularly needs evaluation on real interfaces. Understanding one screenshot and completing a sequence of business actions are different challenges. State changes, scrolling and error messages can alter the next decision.

Who should consider upgrading?

Teams working heavily with product photos, drawings, application screenshots or text editing—and existing GPT-5.6 Terra users—have a reason to run a controlled comparison. Confirmed lower rates justify a cost experiment; reduced correction and retry effort must be established on their own workloads.

For the hardest coding, research or long-running professional reasoning, compare Sol and Astra according to the official model roles rather than selecting Terra on token price alone. An API release also does not establish availability in every ChatGPT subscription's model picker. This article confirms the API documentation's release and specifications.

TboxIO's assessment: the clear attractions are the visual, computer-use and writing focus, and lower published token rates than GPT-5.6 Terra. Current evidence supports a controlled upgrade evaluation, not claims that it is universally stronger, faster or cheaper per task.

Sources

• Official API changelog: October 26 release
• GPT-6.1 Terra model page: specifications, pricing and tools
• Official pricing: processing tiers, caching and tools
• GPT-6 guide: Terra positioning, effort and migration
• GPT-5.6 Terra model page: prior specifications and rates