GPT-6.1 Terra gives teams a cheaper model to test for work that involves looking at screens, interpreting images and acting through software. OpenAI released it on 26 October 2026 with standard API prices of $1 per million input tokens and $5 per million output tokens. The important buying question is whether those lower rates survive a complete workflow: reasoning, tools, retries and the cost of checking the result.
The strongest reason to evaluate Terra is the combination of visual work and routine execution. A model that can identify the right field on an unfamiliar screen has a different economic value from one that merely produces a polished answer. Yet a successful demonstration of entering an invoice is not permission to let the model approve payments. Capability, task suitability and authority remain separate decisions.
This guide uses OpenAI's release announcement, direct model documentation and system card, checked on 28 October. We have not run an independent Terra benchmark. The examples below are calculations and proposed evaluation designs, with vendor results labelled as such.
- 01Start with visual and routine agent work.The release targets screenshots, computer use and writing. Complex research and coding still warrant comparison with Sol or Astra.
- 02The headline bill is only the starting point.Long prompts, reasoning tokens, cache writes and tools can materially change the cost per completed task.
- 03Use the direct API contract.Tool calling requires Responses. Router variants and floating aliases are separate integration choices.
- 04Evaluate what the agent actually changes.A good answer, a correct tool call and a verified business outcome are three different checkpoints.
01 — WorkloadWhere the new model fits #
OpenAI's launch announcement positions Terra around visual reasoning, computer use and natural writing. Its most interesting applications therefore combine observation with action: finding the matching catalogue item, reading a diagram before editing a drawing, or preparing a draft from information visible in several applications. These are candidates for a trial, not proven capabilities on your organisation's documents.
Choose a workflow whose result can be inspected cheaply. For example, ask an agent to extract a supplier name, invoice date and total into a draft record, with a link back to each source page. A reviewer can compare the fields before the record becomes operational. If the invoice is ambiguous or the supplier identity conflicts with an existing record, a correct outcome may be to stop and flag the conflict.
A poor first trial is an open-ended request to reorganise every customer account. That task combines unclear objectives, broad access and many possible side effects. Even an improvement in model intelligence cannot supply missing acceptance criteria. Start with a bounded unit of work, retain the existing route as a comparator, and count both correct completions and justified abstentions. A model that declines the difficult cases may be useful if the handoff is explicit and the remaining workload is valuable.
Visual extraction
Read a document or screen and return a draft with evidence. Verify the fields before committing them.
Complex engineering
Keep difficult debugging and architectural changes in a comparison with stronger models. Do not infer coding parity from a desktop benchmark.
02 — The billRead the prices as a contract #
The direct model card is the relevant source for the OpenAI API. Its standard rates distinguish uncached input, cached input, output and cache creation. Those categories matter because a repeated agent instruction can be cheap to read after caching, while a fresh or rewritten prefix may incur a different charge. A headline price comparison that assumes every input token receives the cache discount is not a budget.
Consider a hypothetical request with 10,000 uncached input tokens and 2,000 billed output tokens. At the standard short-context rates, the model charge is $0.02: $0.01 for input and $0.01 for output. That is an arithmetic example, not an observed cost for processing an invoice. The real task may require several calls, more reasoning than visible prose, a search tool, or a second attempt after a validation failure.
Store usage against the logical task, not only the request. If a workflow makes four calls, its cost is the sum of those calls plus any tool charges and infrastructure costs. Compare accepted results at the same quality threshold. A cheaper request that leaves more work for a reviewer can lose its advantage; a more expensive request that prevents several failed retries can win. The model's list price is an input to that comparison, not its conclusion.
| Standard price category | USD per million tokens |
|---|---|
| Uncached input | $1 |
| Cached input | $0.05 |
| Output | $5 |
| Cache write, 30-minute retention | $1.25 |
| Cache write, 12-hour retention | $2.50 |
Source: OpenAI's direct Terra model card, checked 28 October 2026. These are short-context Standard rates, before tools and other applicable modifiers.
03 — ThresholdsA long prompt can change the bill #
The context window and the low-price band are different limits. Terra's model card lists a 1,050,000-token context window and a maximum output of 128,000 tokens. However, a prompt above 272,000 input tokens triggers higher rates for the full request: twice the input and cache rates and 1.5 times the output rate. Having room for a document does not mean that processing it costs the same per token as a shorter request.
A hypothetical 270,000-token uncached prompt with 10,000 billed output tokens costs $0.32 at Standard rates. Raising the prompt to 280,000 tokens takes the calculation to $0.635. The extra 10,000 input tokens are not the main cause of the increase; crossing the threshold reprices the request. These examples exclude tools, cache writes and other processing modifiers. They illustrate a pricing discontinuity, not a measured performance difference.
Before stuffing a whole project into the prompt, ask whether every section is necessary for the decision. Retrieval can reduce the input, but it can also omit a crucial exception. Summaries can reduce repetition, but they can lose provenance. Evaluate those strategies on the actual task rather than assuming less context is always better. Log the input size, evidence selected and final answer together so a cost reduction can be checked against correctness.
Set separate alerts for total task spend and requests near the long-context threshold. A per-request token ceiling does not explain whether the task became cheaper or merely required more requests.
04 — BenchmarksSeparate a result from a promise #
OpenAI reports strong computer-use results on Agents' Last Exam, including performance comparable with Astra at a much lower per-token price. That result is evidence from a vendor evaluation. It does not establish equivalence across all coding, research and business tasks. A benchmark name also identifies a particular version and harness; changing the tools, reasoning setting or environment can change the result.
The launch's Terminal-Bench 4.0 result is more concrete: OpenAI reports 50.3% at maximum reasoning effort for Terra, compared with GPT-5.6 Terra's published 21.5%. The useful takeaway is that the provider reports a substantial advance over its earlier Terra model. It is not that every terminal task now has a one-in-two chance of success. Your task distribution, permissions and test suite differ from the benchmark's.
Treat the provider's disclosed limits as part of the evidence. OpenAI still recommends Sol and Astra for the most complex coding tasks. An evaluation should therefore include a deliberate failure case: a task you expect the lower-priced model to escalate. A model that confidently completes the wrong thing can look productive in a throughput report. A model that recognises its boundary can be easier to operate, even when its raw completion count is lower.
Keep model version, benchmark version, effort, harness and evaluator attached to every comparison. A percentage without those conditions is too easy to turn into a claim the source never made.
05 — IntegrationCheck the tool loop before switching #
Terra's direct documentation says tool calling requires the Responses API. That is a meaningful integration detail for an application built around another endpoint. A model identifier change can expose differences in tool-result handling, supported parameters and saved conversation state. Check the request contract in the client you actually use instead of assuming an OpenAI-compatible interface guarantees equivalent behaviour.
The documented reasoning choices include none, low, medium, high, xhigh and max, with medium as the default. Pick a setting deliberately for the trial. Otherwise, a model comparison can silently become a comparison between different amounts of reasoning. Record billed output and completion time alongside the answer, and distinguish visible text from the reasoning tokens the service bills.
Image input does not make Terra a native image, audio or video generator. The direct card lists text output and image understanding, while tools may provide additional capabilities. A workflow that requests an image through a tool involves another service contract and potentially another charge. Test tool availability, result format and error handling on the exact route. A route that returns text successfully but cannot execute the required tool has not passed migration.
| Boundary | What to verify |
|---|---|
| Endpoint | Responses API tool calling works in the deployed client |
| Reasoning | The configured effort is supported and explicitly recorded |
| Images | Input handling, resolution assumptions and extracted evidence |
| Tool outcome | Actual tool result and downstream state, not just assistant narration |
Integration checks proposed by Digital Applied; these are acceptance criteria, not measured Terra results.
06 — Version controlPin the model you meant to buy #
Our 28 October OpenRouter capture includes Terra, Terra batch, Terra Pro and Terra Pro batch routes. The Pro description identifies the same underlying model running a higher-reasoning mode. Identical per-token prices would not imply an identical total bill if that mode produces substantially more billed reasoning. Treat route names as configurations to inspect, not as evidence of four independently released foundation models.
The same capture shows the floating Terra alias targeting GPT-6.1 Terra rather than GPT-5.6 Terra. Floating aliases are convenient for exploration, but they can change the model beneath a production workflow. A stable test result needs the exact model route, provider and configuration. If the application intentionally follows an alias, record the resolved target when possible and decide what level of change triggers a new evaluation.
Do not mix router prices with direct-provider benchmark claims without naming both sources. The route can carry a different provider mix, limit or supported parameter list. A team might legitimately choose a router for operational convenience, yet it still needs to test the route it will use. Keep a successful direct-API trial as useful evidence, then add the route-specific checks rather than assuming the first trial proves both integrations.
Save the exact route and resolved model alongside the prompt version and evaluation result. A floating alias is a deliberate update policy, not a frozen model version.
07 — PermissionsCheaper actions still need boundaries #
An agent's ability to operate software increases the importance of controlling what it may change. Reading a webpage, drafting a reply, sending that reply and changing an account setting should not all inherit the same permission simply because the model can perform them. Define authority at the tool boundary and make external side effects observable.
The Terra system card is useful precisely because its results are not uniformly favourable. Different tests examine different failure modes, including responses to messages that appear to come from other agents. Our analysis of the peer-message experiment explains its small sample and grading limits. It should inform testing, not become a claim about a production incident rate.
For an invoice workflow, a sensible first deployment might allow extraction and draft creation while keeping supplier-bank changes and payment approval outside the agent's authority. If a document tells the agent to send information elsewhere, that text is part of the document, not a new instruction from the user. Test this boundary with realistic documents and inspect tool logs. Prompt wording alone is not a substitute for restricting access and verifying outcomes. Permission to inspect information does not automatically include permission to disclose or change it. Keep the permitted actions explicit when a visual workflow crosses applications.
08 — AdoptionRun a trial that can change your mind #
Select a small, representative task set with ordinary examples, ambiguous examples and cases that require stopping. Keep the existing model's settings and Terra's settings in the record. Review the outputs without letting the model label determine the verdict where practical. Decide beforehand which errors are unacceptable and which can be corrected cheaply by a person. Measure accepted completion, reviewer time, total billed usage, elapsed time and unintended actions. Include failures rather than quietly rerunning them until they pass. If a task succeeds on the third attempt, its time and cost include all three. A pilot that reports only successful requests will tend to exaggerate the benefit of any model with inconsistent behaviour.
Finally, make the migration reversible. Start with a bounded portion of eligible work, retain an escalation route and keep the prior configuration available. Move more work only when the observed result justifies it. Terra's release is a strong reason to revisit the economics of visual and routine agent tasks; it is not a reason to discard the controls that made the workflow dependable. The useful outcome is a lower cost per accepted task under the same authority and quality requirements.
Adopt when the workflow meets its quality and permission requirements at a better total cost. Keep a stronger model or a human handoff for the cases that fail that test.
Continue with the existing Sol upgrade analysis, long-context pricing, computer-use architecture, the peer-message safety experiment. For implementation support, explore our AI transformation services.
Evaluate the whole workflow before changing the default.
Terra makes visual and routine agent work worth testing again. Keep the price conditions, route contract and evidence limits visible, then judge the model by accepted outcomes and the effort needed to verify them. A cheaper token becomes a useful saving only when the surrounding workflow still works.