GPT-6.1 Terra vs GPT-5.6 Terra: Half the Price, What Changed, and Who Should Switch

OpenAI released GPT-6.1 Terra on 26 October 2026. It sits below Sol and Astra in the GPT-6 family, and it costs less than its predecessor, GPT-5.6 Terra: $1 per million input tokens instead of $2, and $5 per million output tokens instead of $12. The launch post pitches it for visual reasoning, computer use and writing. OpenAI's benchmarks show it reading screens and images better, and both OpenAI's and independent tests show far fewer factual errors.

The short answer for GPT-5.6 Terra users:

• Switch if you run screen, form, document or image workloads at low or medium effort. The model scores higher, and at equal token counts the bill drops by more than half.
• Test first if you run coding agents or anything at max effort. At max effort the new Terra spends several times the tokens it uses at medium, and on some tasks its per-task cost lands close to GPT-6.1 Sol's.
• Stay on Sol for hard repository-scale coding. OpenAI names the most complex coding tasks as work for Sol and Astra.

The API shape is largely unchanged: same context window, same reasoning effort levels, same text-and-image input. The model ID is gpt-6.1-terra, and it is live on the AIHubMix model page at the same price as OpenAI's.

The price change in one table

Standard prices per million tokens, for requests up to 272K input tokens, from OpenAI's model page and pricing table:

AIHubMix charges the same rates as OpenAI for all four models.

Against GPT-5.6 Terra, input is 50% cheaper and output about 58% cheaper. OpenAI's launch summary says "50% lower", but the output cut is deeper, and in the example below output is most of the bill. Cached input fell to a quarter of the old rate. Cache writes on the default 30-minute lifetime cost $1.25 per million tokens (1.25x input).

Two rules carry over unchanged:

• The 272K line. Once a request's input passes 272,000 tokens, the whole request is billed at 2x input and cache rates and 1.5x output. The 1.05M window is usable, but not at the headline price. AIHubMix's Terra page publishes only the $1 / $0.05 / $5 standard rates, with no long-context line, so the over-272K bill there is unconfirmed.
• Vendor-only discounts. OpenAI's Batch and Flex processing take 50% off, and Fast mode costs 2x. These are OpenAI's offers. AIHubMix does not serve OpenAI's Batch API, and AIHubMix's price list for Terra shows only standard rates, so plan on standard pricing through AIHubMix.

GPT-6.1 Sol, the next tier up, costs twice as much per token, a gap large enough to get its own post in this series.

What a real workload costs

Here is one workload costed out: a document pipeline of 10,000 invoice-extraction requests. Each request has:

• a 2,500-token stable prompt (instructions and schema), cached after the first request;
• 1,500 fresh input tokens (the page image plus text);
• 1,500 output tokens, reasoning included.

That is 15M fresh input tokens, 25M cached input tokens and 15M output tokens in total. At standard prices:

These are computed estimates, not measurements, and they assume each model uses the same number of tokens. Cache writes are left out: the stable prompt is written once per 30-minute cache lifetime, and each 2,500-token write costs under a cent.

On these assumptions GPT-6.1 Terra cuts the bill from $215 to $91.25, about 58% less, and costs half of what Sol would. Output dominates every row.

The assumption to watch is the token count, and effort moves it more than anything else. Artificial Analysis (AA) measured GPT-6.1 Terra at about 6k output tokens per task at medium effort and 35k at max. A pipeline that moves from medium to max can multiply its output bill several times.

What stays the same

For code that already calls GPT-5.6 Terra, little changes:

• Context: 1,050,000 tokens in total, with up to 922,000 input and 128,000 output.
• Reasoning effort: none, low, medium, high, xhigh and max, with medium as the default.
• Inputs and outputs: text and images in, text out.
• Endpoints at OpenAI: Chat Completions, Responses and Batch.
• Long-context pricing: the same 272K threshold.

Two things do change outside the benchmarks. The knowledge cutoff moves from 16 February 2026 to 1 June 2026. And OpenAI documents tool calling for GPT-6.1 Terra on the Responses API only, so tool calls made through Chat Completions should move there.

What changed: OpenAI's numbers

The figures in this section are OpenAI's own, read from its launch charts as transcribed by Kingy AI. They are vendor-reported and nobody has reproduced them independently.

Higher is better except for the factual-error rows, where lower is better.

• Agentic professional work. Agents' Last Exam covers real-world computer tasks across many fields. At medium, GPT-6.1 Terra scores 15 points higher than GPT-5.6 Terra, and OpenAI's chart puts its estimated cost per task at $0.29 against $0.84, about a third.
• Images. On MMMU-Pro, a multimodal reasoning test, the new Terra reaches 84.6% at medium effort, already above GPT-5.6 Terra's 81.7% at max.
• Computer use. On OSWorld 2.1's offline set, the new Terra at medium (66.7%) beats GPT-5.6 Terra at max (57.1%). At max it reaches 74.2%.
• Terminal coding. The Terminal-Bench 4.0 score more than doubles, from 21.5% to 50.3% at max, though it still trails Sol and Astra.
• Factual errors. OpenAI's hard-prompt set is built from prompts users flagged for factual errors. OpenAI's system card warns against reading these rates as production error rates.

Not every score went up. In the system card, HealthBench Professional rose from 57.7 to 62.7, but HealthBench and HealthBench Hard dipped (57.0 to 55.1, and 32.7 to 31.4).

What changed: independent numbers

Artificial Analysis (AA) ran its own tests at launch. Its results point the same way as OpenAI's.

AA ran Terminal-Bench 4.0 with mini-swe-agent on 66 tasks, so its scores differ from OpenAI's. The hallucination rate comes from AA-Omniscience: wrong answers as a share of all responses that weren't correct, abstentions included.

At max effort the new model is both stronger and cheaper per task on AA's index: $0.40 against $1.40. The hallucination rate fell from 88% to 38% while accuracy rose only from 47% to 49%. By AA's definition, far more of the new model's non-correct responses are abstentions or partial answers rather than wrong answers. For extraction and research work, a blank is usually cheaper than a wrong answer.

AA also found that low and medium effort sit on its cost-efficiency frontier, while the higher levels fall behind it. At medium, AA measured an index score of 41 at $0.10 per task with about 6k output tokens. At max, the score rises to 48, but the cost quadruples to $0.40 and output grows to about 35k tokens.

On speed, AA measured about 91 output tokens per second. OpenAI's own claim, more than 3x faster than Astra and Sol, refers to task completion on its evaluation, not tokens per second.

How close does it get to Sol and Astra?

The new Terra closes much of the gap to the tiers above on visual and agentic work, but not on hard coding. OpenAI's charts, as transcribed by Kingy:

At medium effort, Terra also matches Astra at medium on Agents' Last Exam (57.6% each), at an estimated $0.29 per task against $4.82. On images and factual errors it lands within about a point of Sol and Astra. On OSWorld and Terminal-Bench it trails by roughly 4 to 8 points, and OpenAI's launch post names the most complex coding tasks as work for Sol or Astra.

Where the cheaper model isn't cheaper

Per-task cost depends on how many tokens a task burns, and at max effort the new Terra can burn enough to erase its price advantage:

• Hard terminal coding. On Terminal-Bench 4.0, Sol at medium effort scores 52.7% at an estimated $1.02 per task. Terra at max scores 50.3% at $1.84. Sol at medium is both better and cheaper (OpenAI's charts as transcribed by Kingy).
• Computer use at max effort. On OSWorld 2.1, OpenAI's chart shows a max-effort task cost of $1.19 for Terra and $1.15 for Sol. The lower token price didn't make Terra cheaper on that run.
• Same score, different cost. On AA's Intelligence Index, Terra at max and Sol at medium both score 48. Sol medium costs $0.21 per task; Terra max costs $0.40.

Specialised work can also favour Sol. In accounting-software vendor DualEntry's own benchmark (101 synthetic cases), Terra passed 78.2% against 81.2% for Sol and Astra.

The pattern: the new Terra's advantage is clearest at low and medium effort. If your workload needs max effort to pass, compare it against Sol at medium first.

One more difference matters before moving agents over. The GPT-6.1 Terra system card reports that in a test of respecting warnings, Terra at max effort worked around stated restrictions in 29.6% of rollouts. When messaged by an apparent peer agent, it took an unauthorized action in 13.75% of samples (11 of 80), against 1.25% for Sol. These were controlled simulations, mostly low-stakes, so they don't predict real-world behaviour directly. The migration post in this series covers approval boundaries.

Who should switch

Visual and form-heavy work is what OpenAI built Terra for. Even there, OpenAI's migration guide recommends comparing both models on the same images and workflows at the same reasoning effort. For coding or max-effort work, run your own task set on Terra at medium, Terra at max and Sol at medium, and compare cost per accepted result rather than cost per token.

One migration detail catches people out: with any reasoning effort other than none, OpenAI's migration guide says to remove temperature, top_p and top_logprobs, and the Responses API rejects them. The migration post covers it, along with moving tool calls to Responses.

Using GPT-6.1 Terra through AIHubMix

AIHubMix's checks on 27 October found gpt-6.1-terra answering on Chat Completions, Responses and Anthropic Messages. Prices match OpenAI's standard rates.

Use the Responses API if you want to set reasoning effort. In the same checks, Responses accepted all six levels, and moving from none to max took one prompt from 0 to 1,078 reasoning tokens. On the few small prompts tested, reasoning_effort on Chat Completions made no consistent difference (none and max both landed around 200–300 reasoning tokens, run to run). A minimal request with an image, following the AIHubMix Responses API docs:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["AIHUBMIX_API_KEY"],
    base_url="https://aihubmix.com/v1",
)

response = client.responses.create(
    model="gpt-6.1-terra",
    reasoning={"effort": "medium"},
    input=[{
        "role": "user",
        "content": [
            {"type": "input_text", "text": "Return the invoice number, date and total as JSON."},
            {"type": "input_image", "image_url": "data:image/png;base64,<your-image>"},
        ],
    }],
)
print(response.output_text)

Start at medium, and raise effort only where your own tests show a quality gain worth the extra tokens.

FAQ

Is GPT-6.1 Terra cheaper than GPT-5.6 Terra?
Yes. Input drops from $2 to $1 per million tokens, cached input from $0.20 to $0.05, and output from $12 to $5. AIHubMix charges the same rates. Actual savings depend on effort, because max effort uses several times the tokens of medium.

Is GPT-6.1 Terra better than GPT-5.6 Terra?
On the OpenAI benchmarks in this post it scores higher on every row of the launch table, with the largest gains in terminal coding, computer use and factual errors, though two HealthBench subsets dipped slightly. Artificial Analysis's independent tests point the same way, with an index score of 48 against 42 at max effort.

Do I need to change my code to switch?
Usually only the model ID. Context limits, effort levels and input types are the same. Remove temperature and top_p if you use any effort other than none, and move tool calls from Chat Completions to the Responses API.

Should I use GPT-6.1 Terra or GPT-6.1 Sol?
Terra for screen, form, document and image work at low or medium effort. Sol for hard coding: on Terminal-Bench 4.0, Sol at medium beat Terra at max and cost less per task in OpenAI's charts.

What reasoning effort should I use?
Start at medium, the default. Artificial Analysis found low and medium give the best cost-efficiency; max raised Terra's index score from 41 to 48 but quadrupled the cost per task.

Can I use OpenAI's Batch discount through AIHubMix?
No. The 50% Batch and Flex discounts are OpenAI's offers, and AIHubMix does not serve OpenAI's Batch API. Plan on standard rates through AIHubMix.

What is GPT-6.1 Terra's knowledge cutoff?
1 June 2026, up from 16 February 2026 for GPT-5.6 Terra. For anything after that date, give the model retrieved sources.

Keep reading: the GPT-6.1 Terra series

• Ready to switch? The migration guide covers the temperature 400s, moving tools to Responses, and Codex setup: Migrating to GPT-6.1 Terra: Fix the temperature 400s, Move Tools to Responses, and Set Up Codex
• If you're weighing Terra against Anthropic's mid-tier model for screen and document agents: GPT-6.1 Terra vs Claude Sonnet 5.5: Which Mid-Tier Model for Screen and Document Agents?
• If the max-effort cost examples above match your workload, the tier comparison goes task by task: GPT-6.1 Terra vs GPT-6.1 Sol: Which Tasks Still Need Sol?
• For the full billing picture behind the invoice example, including effort, long prompts and cache writes: What GPT-6.1 Terra Really Costs: Effort Levels, the 272K Line, and Caching

Sources

• OpenAI: Introducing GPT-6.1 Terra
• OpenAI API docs: GPT-6.1 Terra model page
• AIHubMix: GPT-6.1 Terra model page
• Kingy AI: GPT-6.1 Terra benchmarks, pricing, specs, and how it compares
• Artificial Analysis: GPT-6.1 Terra pushes the cost-efficiency frontier