GPT-6.1 Terra: What Changed, What It Costs, When to Use It

OpenAI introduced GPT-6.1 Terra as the cheap, fast member of the GPT-6 family: visual reasoning, computer use, clearer writing, and API prices half of GPT-5.6 Terra and GPT-6.1 Sol. Founders will see the model id appear in Codex, ChatGPT Work, Lovable app features, and any agent that auto-picks "latest." A sticker-price cut is not a reason to change a production call. A model release still changes three things you pay for: tokens, the behavior of prompts you already tuned, and — if the API surface moved — whether the call runs at all.

This note is for apps that already call OpenAI, not for people ranking models for fun. We cover what Terra is, what changed from 5.6 Terra, the migration traps OpenAI documents, how effort and long context change the bill, and the decision rule for whether your path should move. The same rule we used for Claude Fable 5.2 applies here: evals first, then price. The method for turning that into a monthly number is in AI app maintenance cost.

What GPT-6.1 Terra actually is

OpenAI's announcement describes Terra as a significant step up from GPT-5.6 Terra on the work that fills a workday: drafting, reading drawings, filling forms, editing video timelines. It is not the flagship. GPT-6 Astra is the demanding-work model. GPT-6.1 Sol is complex work at lower cost than Astra. GPT-6 Luna is high-volume focused tasks. The GPT-6 guide puts Terra in the visual-reasoning-and-writing slot.

The numbers that matter in an app, from the model page:

• Model ID: gpt-6.1-terra (Bedrock: openai.gpt-6.1-terra)
• Context: 1,050,000 tokens; maximum input 922,000; maximum output 128,000
• Modalities: text and image in; text out. Audio and video are unsupported
• Knowledge cutoff: 1 June 2026
• Reasoning effort: none, low, medium (default), high, xhigh, max
• Endpoints: Chat Completions and Responses, plus Batch. Live, Realtime, Assistants, fine-tuning, embeddings, image and video generation, and speech are not this model's job
• Tools (Responses API): web search, file search, image generation, code interpreter, hosted shell, apply_patch, skills, computer use, MCP, tool search
• Availability: Plus, Pro, Enterprise, and Edu in ChatGPT Work and Codex; developers on the API. Lovable listed Terra for in-app reasoning and image analysis on 6 November 2026

OpenAI's own coding line is the one to keep: Terminal-Bench 4.0 at maximum effort, Terra solves 50.3% of tasks, more than double 5.6 Terra's 21.5%. For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice. A leaderboard row that looks strong on cost is not a reason to put Terra on a multi-hour agent path you already run on Astra.

Safety: Terra meets the Critical cybersecurity threshold in OpenAI's Preparedness Framework and keeps Astra-era protections. The announcement claims better scope-staying, more honest limitation disclosure versus 5.6 Terra, and stronger prompt-injection resistance in their tests. Those are vendor evals. They are not your product's evals.

What changed from GPT-5.6 Terra

Price dropped by half on the sticker. Standard API is $1 / million input, $0.05 / million cached input, $5 / million output. OpenAI states that is 50% below GPT-5.6 Terra and GPT-6.1 Sol. Cache writes are billed separately: $1.25 / million for 30-minute retention, $2.50 for 12-hour, before long-context adjustments. Cached reads are 5% of uncached input. Batch and Flex are 50% below Standard. Regional processing adds 10% where offered.

Fast mode is a different product. Fast mode is 2× Standard. Same model family, double the token rates. Use it when latency is the feature. Do not leave it on a nightly invoice job.

Effort is a first-class control, and Terra still allows none. Astra and 6.1 Sol do not support none; Terra, GPT-6 Sol, and Luna do. Default is medium. OpenAI's migration note: if you used minimal on an older model, start at low and compare. Harder effort means more reasoning tokens before the answer you bill as output. Cheaper rates times more tokens is not automatically cheaper.

Factuality and coding scores moved. On OpenAI's internal error-inducing prompts, low-effort error rate fell from 53.6% (5.6 Terra) to 16.7% (6.1 Terra). At max effort the rate fell 18 points to 4.3%, near Astra and Sol. Visual benchmarks (MMMU-Pro, OSWorld 2.1, Agents' Last Exam) are where OpenAI spends the rest of the announcement: Terra closing on Astra at a fraction of the cost, and faster than Astra and Sol on those computer-use tasks. That is the workload the model is for. Invoice OCR and catalog matching are in-scope. Ambiguous multi-day coding is the Sol/Astra sentence.

Independent coding benches are less flattering at the cheap effort levels. AI Coding Daily, last run 5 November 2026 on Codex CLI, put Terra Max at 77.11 / 90 and about $0.27 per prompt, versus GPT-6 Astra High at 84.82 and $0.96. Terra Medium sat at 68.5. Effort is not a free upgrade; Max was the slow configuration on that board (about eight minutes average). A default-medium switch "because it is cheaper" can look like a quality regression on code review.

What can break when you change the model id

OpenAI published a migration quickstart. These are the items that fail in production if you only swap the string.

1. Tool calling requires the Responses API. Chat Completions is supported for GPT-6.1 Terra. Function calling is not, on that endpoint. Agents that still POST tools to /v1/chat/completions need to move to Responses before the upgrade. GPT-6 Sol and Luna still allow Chat Completions function calling only with reasoning_effort: "none". Terra is in the "use Responses for tools" group with Astra and 6.1 Sol.

2. Sampling parameters are invalid when effort is not none. Remove temperature, top_p, and top_logprobs. On Chat Completions also remove logprobs. On Responses, remove message.output_text.logprobs from include. A 400 here is not a "the model is worse" event. It is a leftover parameter from a 5.x call.

3. Prompt cache fields renamed if you came from GPT-5.5 or earlier. Replace prompt_cache_retention with prompt_cache_options.ttl set to "30m". Review cache boundaries and cache-write billing. A silent cache miss shows up as a higher input bill, the same way Fable 5.2's prefix check did.

4. Long prompts are a different price, for the whole request. Inputs above 272K tokens bill at 2× input and cache and 1.5× output for the full request. A RAG path that stuffs a repo into context can erase the 50% sticker cut in one call. Log input tokens per feature before you celebrate the price list.

5. Compare at the same effort, on your images. OpenAI's Terra-to-Terra note is explicit: when moving from 5.6 Terra to 6.1 Terra, compare both on representative images and workflows at the same reasoning effort. A default of medium on 6.1 versus whatever 5.6 was using is not a controlled upgrade.

6. Fast mode and data residency. Fast mode supports EU residency for Terra (and 6.1 Sol and Luna), not for Astra. If your app pinned a region, check the fast-mode table before you flip a latency flag.

None of these are reasons to avoid Terra. They are reasons to upgrade behind a flag, with a rollback, rather than because Codex or Lovable offered the new id in a dropdown.

What the new price does to the bill

The maintenance-cost post's rule still holds: monthly model spend is features × successful tasks, not users. Terra changes several variables at once.

• Sticker input and output are half of 5.6 Terra on Standard.
• Default effort is medium with reasoning tokens. Output can rise.
• Fast mode doubles both sides.
• Cache writes are not free.
• 272K+ prompts reprice the entire request.
• Computer-use and other tools have per-call fees on top of tokens — the model page points at the live pricing page rather than a number to memorize.

The honest metric is cost per successful task at the effort you actually run, including retries. OpenAI's own computer-use story is that Terra at medium effort scored 15 points above 5.6 Terra on Agents' Last Exam at about one-third the cost, and on par with Astra at about one-tenth the price per token. That is their bench. Your invoice OCR pipeline may match it. Your "fix this flaky test suite" agent may not. Measure on your logs.

For ChatGPT Work and Codex subscribers, OpenAI says the same plan allowance covers more Terra tasks than 5.6 Terra, and substantially more than Sol or Astra. That is a product-plan claim, not an API invoice. If the app bills you per token, ignore the Work copy.

Terra, Sol, Astra, or stay: choose with evals

OpenAI's family split is usable as a decision rule:

1. Luna — focused, high-volume, cheapest-shaped work.
2. Terra — visual reasoning, computer use, writing, high-volume automation (invoices, catalogs, drawings) where 5.6 Terra was already the model.
3. 6.1 Sol — complex work, including harder coding, without Astra's price.
4. Astra — the demanding remainder.

OpenAI's coding sentence is the override: most complex coding stays on Sol or Astra. AI Coding Daily's spread between Terra Medium and Astra High is the same advice with different numbers.

Do not switch because Lovable, Cursor, or Codex offered Terra in a picker. Do this:

1. Write down 5–20 real user inputs (include images if the feature is visual).
2. Record what the current model returns and what "correct" means.
3. Run them on gpt-6.1-terra at the same effort, then at default medium if that differs.
4. Keep the change only if the failures you care about get fewer and cost per success is better, retries included.
5. Ship behind a flag. Pin the snapshot id.

That small suite is what a Production MVP includes, because model ids will keep moving on days you did not plan.

If the app already calls OpenAI: a migration checklist

1. Read the GPT-6 guide against your client. Grep for Chat Completions + tools, temperature, prompt_cache_retention, minimal effort, and hardcoded gpt-5.6-terra.
2. Move tool paths to Responses before you change the model id.
3. Drop illegal sampling params when effort is not none.
4. Cap and log input size. Anything near 272K needs a chunking or retrieval design, not a bigger window.
5. Run the eval set. Images and computer-use traces if that is the feature; code tasks if you were tempted to replace Sol.
6. Measure cost per success for a few days on a shadow or canary, Fast mode as a separate configuration.
7. Watch the first week. Behavioral drift is quiet: a prompt that relied on 5.6 Terra's verbosity, or an agent that now asks for confirmation more than your UI expects. The GPT-6 prompting notes on initiative and follow-through are written around Astra; still test Terra against "it stopped to ask" versus "it assumed."

Lovable apps that gained Terra on 6 November for "reasoning and image analysis" still need the same loop. A builder dropdown is not an eval.

Where Terra belongs in a production app — and where it does not

The announcement is strongest on computer use and vision. Agents' Last Exam, OSWorld, MMMU-Pro, and the invoice example are that cluster: look at a screen or a document, act, write a log. If your feature is "match this photo to a SKU," "extract line items," or "fill this form from a PDF," Terra is the model OpenAI is asking you to try, at half of 5.6 Terra, with none still available when you need a cheap non-reasoning turn.

It is weaker as a default coding agent. OpenAI says so. The independent Codex CLI board says so at medium effort. Putting Terra on every Cursor or Claude-Code-shaped path because the unit price fell is how you buy a quieter, cheaper model that fails the tasks you actually ship. Keep Sol or Astra (or Fable, or Opus) on the path whose eval already passes, and try Terra on the visual or high-volume side path.

Computer use and MCP are supported tools on Responses, not a reason to point a cloud agent at production credentials. The PocketOS outage was a token that could delete a volume. Terra being better at OSWorld does not change that. Hosted shell and computer use in an app that holds customer data still need a token that cannot destroy production and a person in front of irreversible calls.

Context window is not free context. A million tokens of window with a 272K price cliff means retrieval design still matters. Stuffing a repo into the prompt to "use the new context" is how the 50% cut disappears and latency returns. Prompt cache helps only if the prefix is stable; the migration guide's configuration_update trick exists so you can change effort without busting that prefix. Use it if you already cache. Do not invent a 900K prompt to justify the window.

Unsupported surfaces are not bugs. No realtime, no native audio out, no fine-tune, no Assistants API. If the app is a voice agent or a fine-tuned 4.x classifier, Terra is the wrong row on the pricing page. Check the endpoint table on the model doc before you copy the id into an old Assistants call.

How Techtribe helps you absorb the next OpenAI release

On the scope call we look at how the app calls models: where the id lives, Chat Completions versus Responses, whether tools will 400 on Terra, what a task costs today, and whether you have five inputs you trust. You leave with the short list and a fixed price.

When the product needs routing, fallbacks, evals, and cost per feature, that is a Production MVP, seven thousand dollars, four to six weeks. When model work keeps arriving, the Dedicated Team carries it. We build on OpenAI, Claude, and other providers as a normal week. Prices are on packages.

Bring the repo, one failing or expensive call, and the model id you are on. If the only change you want is gpt-6.1-terra in an env file, we will tell you in ten minutes whether the tools endpoint will survive it.