GPT-6.1 Terra Guide: Setup, Computer Use & Costs

Your invoice bot just got cheaper overnight

Same agent loop: screenshot the accounting UI, read the PDF, type the line items. On GPT-6.1 Sol the token bill climbs fast. GPT-6.1 Terra shipped October 26, 2026 – model ID gpt-6.1-terra – for high-volume screen and document work at $1 / $5 per million input / output tokens on the standard API.

Not a launch recap. Paths only: Work/Codex picker, Responses API starter, effort dials (including none), Free-tier exclusions, and the 272K pricing trap most day-one posts skip.

What GPT-6.1 Terra is (and isn’t)

$1.00 input, $0.05 cached input, $5.00 output per 1M tokens – plus cache writes at $1.25 (30-min) or $2.50 (12-hour). That stack is on the OpenAI model page. Context window: 1,050,000 tokens. Max output: 128,000. Knowledge cutoff: June 1, 2026. Text and image in, text out. No native audio or video.

OpenAI’s own Agents’ Last Exam numbers (medium effort): roughly 15 points above GPT-5.6 Terra at about one-third the cost per task; on par with GPT-6 Astra at ~1/10 the per-token price; wall clock over 3× faster than Astra and GPT-6.1 Sol. Terminal-Bench 4.0 at max effort: 50.3% vs the prior Terra’s 21.5%. Internal error-inducing prompts: low-effort errors 53.6% → 16.7%; max-effort 4.3%. Full write-up lives in the Introducing GPT-6.1 Terra post.

Artificial Analysis puts low/medium effort on the Pareto cost-efficiency frontier and shows guess-rate-when-unknown on AA-Omniscience falling 88% → 38%. Sol and Astra still take the hardest coding and research jobs per OpenAI. Practical split: Terra for volume computer use and clean writing; Sol/Astra when the last reliability points beat burn rate.

Practical setup: ChatGPT Work, Codex, and the API

Regular Chat does not get Terra. Plus, Pro, Enterprise, and Edu get it inside ChatGPT Work and Codex. Free/Go Chat – no. API Free tier – not supported. Enterprise/Edu admins sometimes have to flip the model on.

1. Open the ChatGPT desktop app or ChatGPT Work on the web (mobile Work works too).
2. Under the composer, open the model control. Default / Power is fine as a starting point; open Advanced.
3. Select GPT-6.1 Terra. Reasoning level: medium is the default that usually holds.
4. Codex CLI: pick Terra in the model picker, or pass it with /model / -m.
5. API: model="gpt-6.1-terra" on the Responses API when you need tools. Also GA on Amazon Bedrock and GitHub Copilot (Pro and up), per AWS and GitHub launch notes tied to the same rollout.

Same Work plan allowance goes further on Terra than on GPT-5.6 Terra and much further than Sol/Astra – OpenAI’s framing. Measure your workspace. As of launch day (Oct 26, 2026), independent write-ups still couldn’t find Terra’s exact Business/Enterprise credit burn on the public rate card. That may have changed; check the card before you forecast seats.

If your invoice bot only runs a few dozen docs a day, the price cut is nice and quiet. If it runs thousands, the difference is the line item that decides whether the agent stays on overnight.

First API call and computer-use path

Plain text/image can ride Chat Completions. Tools – computer use, web search, MCP, the rest – need Responses. Minimal medium-effort call:

from openai import OpenAI
client = OpenAI()

resp = client.responses.create(
 model="gpt-6.1-terra",
 reasoning={"effort": "medium"},
 input=(
 "You are processing vendor invoices. "
 "Extract vendor, date, total, and line items as JSON. "
 "If a field is unreadable, set it to null and flag it."
 ),
)
print(resp.output_text)

Screen control: turn on the computer tool in Responses, then loop screenshot → model returns clicks/types → your driver runs them → send the new screenshot back. Official computer-use docs use a Playwright-style sandbox pattern – swap the model string to gpt-6.1-terra, cap max steps so a stuck UI doesn’t print money. Medium effort first on form-fill and catalog matching. Raise effort only when your eval set shows a real quality jump.

Pro tip:reasoning.effort: "none" works on Terra. Community notes on the OpenAI Developer forum thread call out the side effect: temperature and top_p enable. Those stay blocked on gpt-6.1-sol. Use none for pure extraction or style-sensitive drafting where sampling beats deep planning.

Dialing effort, cache, and the 272K trap

Effort on Terra: none, low, medium (default), high, xhigh, max. Start medium; raise only if quality gains pay for the extra time and tokens. Cached input at $0.05/M is 5% of uncached – agents that re-read the same system prompt or style guide should cache that prefix hard.

The catch is the long-context row:

Cross 272K input tokens and – per the model page – 2× input/cache and 1.5× output apply to the entire request, not the overflow slice. A casual 400K context can erase the headline discount. Regional processing adds 10% where offered. Batch/Flex sits at half standard; Fast doubles it.

Speed demos look almost unfair until you remember the training target: short, decisive computer steps – not Astra-length research monologues. Run your own traces. Personality fit won’t show up on a launch chart.

Honest limits before you migrate everything

Hardest coding and long research? OpenAI still points you at GPT-6.1 Sol or GPT-6 Astra. Terra’s Terminal-Bench jump is real; the company still routes the nasty jobs upward. Audio and video aren’t modalities here. Safety tier: Critical cybersecurity under the Preparedness Framework, Astra-class safeguard stack – read OpenAI’s system card addendum for the rollout before you ship agentic computer use in production.

Still firing tools through Chat Completions? Those calls fail on this model. Move them to Responses. Compare Terra vs your current model on the same images and workflows at the same effort before you flip production traffic.

FAQ

Is GPT-6.1 Terra in normal ChatGPT chat?

No. Work and Codex only (Plus and above), plus API, Copilot, and Bedrock.

Should I replace Sol with Terra for coding agents?

Scoped terminal tasks, form fills, screenshot office flows, high-volume visual matching – test Terra first. Cost and speed favor it on that band. Deep multi-file refactors, fuzzy architecture calls, long research threads – keep Sol or Astra. Freeze an eval pack: medium effort, both models, same cases. Compare pass rate, tokens, wall time. Cut over only after that sheet says so.

What’s the biggest surprise cost?

People budget the $1/$5 headline and forget two multipliers. One: input past 272K reprices the whole request (2×/1.5×). Two: output-heavy agents live on the $5/M line. Anything that can wait goes Batch/Flex. Repeated prefixes go to cache. Pull one real day of traffic splits before you forecast.

Next step: Work or Codex → Advanced → GPT-6.1 Terra at medium → rerun one invoice or form job you already trust. Log tokens and time against Sol. One comparison beats another chart.