GPT-6.1 Terra Costs $1/M Input: What Should Move to It

OpenAI released GPT-6.1 Terra on October 26, 2026 at $1 per million input tokens — half the price of the model it replaces.

Terra is not a flagship. It is the mid-tier model that clicks buttons, reads forms and turns drawings into documents, and OpenAI just cut what that work costs per token. If you run agents against real screens at volume, this is the line item that moves.

What Terra actually is

Terra replaces GPT-5.6 Terra in OpenAI's lineup. It sits above GPT-6 Luna, the cheapest model, and below GPT-6.1 Sol and the flagship GPT-6 Astra. It is a reasoning model built for high-volume visual and computer-use work: it takes text and image input, carries a 1.05-million-token context window with up to 128,000 output tokens, and can operate a computer — look at a screen, click, type and move between applications.

Access is paid-plans-and-API. It is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.1-terra. Amazon Web Services made it available on Bedrock the same day. It is not in regular Chat, and no free tier is listed.

Through the Responses API it supports computer use, web search, file search, code interpreter, hosted shell and MCP. OpenAI's own example workloads name the buyer precisely: drafting proposals, interpreting technical drawings, filling out forms, entering invoices into accounting software, matching product photos to catalogue entries, and editing video by scene.

The price is the story

Standard API pricing, per million tokens:

Model	Input	Cached input	Output
GPT-6.1 Terra	$1	$0.05	$5
GPT-5.6 Terra (replaced)	$2	—	$12
GPT-6.1 Sol	$2	—	$10
GPT-6 Astra (flagship)	$10	—	$50
GPT-6 Luna	$0.10	—	$0.50

Cached-input rates for the other four models were not listed in the launch pricing used here. Terra lands at half the price of both its predecessor and Sol, and roughly one-tenth the token price of Astra. Luna is still about a tenth of Terra.

Two pricing details that decide your real bill

The long-prompt cliff. Prompts longer than 272,000 input tokens are billed at 2x the input and cache rates and 1.5x output — for the whole request, not just the part over the line. Feed Terra very long documents and that cliff, not the headline rate, is your price.

Batch and Flex. Both are 50% cheaper than Standard; Fast mode costs 2x Standard. Overnight, non-urgent agent batches halve the bill again on the Batch/Flex route.

Those two facts change the shape of the decision. A $1 headline rate on a workload that routinely crosses 272K tokens is not a $1 rate, and a queue-tolerant batch job is not paying $1 either.

What OpenAI claims — and who measured it

Every benchmark figure below is OpenAI's own measurement from its launch announcement. None has been independently verified.

• On Agents' Last Exam — professional tasks such as turning architectural drawings into a 3D model — OpenAI says Terra at medium reasoning effort scores 15 percentage points higher than GPT-5.6 Terra at about one-third the cost, and performs on par with GPT-6 Astra at one-tenth its price per token.
• OpenAI says Terra finishes tasks more than three times faster than Astra or Sol.
• At maximum effort, OpenAI says Terra is over three times as token-efficient as Anthropic's Claude Opus 5.5 at a quarter of its price per token — which works out to roughly a 9x lower cost per task.
• On Terminal-Bench 4.0, a coding test, Terra solved 50.3% of tasks at maximum effort, up from 21.5% for its predecessor.
• On an internal set of prompts designed to trigger errors, the error rate at low reasoning effort fell from 53.6% to 16.7%.
• On OSWorld 2.1, which tests desktop-application use, Terra beat its predecessor's best score at medium reasoning effort while costing less per task.

Vendor benchmarks are a direction, not a budget. Treat all of the above as claims to verify on your own workload.

Where Terra does not belong

Terra is not the leaderboard model. On Terminal-Bench, its 50.3% trails Sol's 55.9% and Claude Sonnet 5.5's 70.6% — figures reported in independent summaries of the launch. OpenAI itself still points complex coding to Sol or Astra.

The split is clean enough to act on. Hard coding and deep research stay on Sol or Astra. Tasks that need no visual reasoning or computer use belong on Luna, which is cheaper still. Terra takes the middle: repetitive, screen-bound, token-billed volume.

Price pressure is not unique to this launch. One developer writing about agent economics this month claims API prices went up on 10 of 14 model tiers in 2026 and that rate limits tightened — their argument for running small local models inside agent loops instead. Take that as one practitioner's read of the market, not a measured industry figure; the point that survives is that per-step agent cost is now a design constraint, not an afterthought.

What to do this week

1. Audit by workload shape, not by model name. Anything screen-bound, form-bound or document-bound currently on Sol or GPT-5.6 Terra is a candidate for gpt-6.1-terra. Hard-coding-bound work is not.
2. Measure your prompt length distribution against 272,000 tokens. If a meaningful share of requests crosses it, chunk them or budget at the higher rate before you promise anyone a 50% saving.
3. Do not give a computer-use model unsupervised write access on day one. Terra clicks and types — that is the capability and the risk. Start on read-only or review-queued work, check what it actually did, then widen scope.
4. Run one independent benchmark on your own tasks before rewriting a model strategy around launch-day charts.

The number to watch next is not another vendor chart. It is the first independent OSWorld-style run on messy, real desktops — that is what tells you whether the price cut survives contact with production.