GPT-6.1 Terra: A One-Line Model Swap, a 58% Cheaper Bill, and Three Things That Break

GPT-6.1 Terra went live on 26 October 2026 at $1 per million input tokens and $5 per million output, against $2 and $12 for GPT-5.6 Terra. For an existing integration the code change is usually the model string, gpt-6.1-terra: the context window, the effort ladder and the input types are unchanged. What does change is the bill, plus three details that either fail loudly at request time or cost money quietly in production.

What is actually identical

Two things outside the table are not neutral. The knowledge cutoff moved forward, so anything you grounded with retrieval for dates between February and June 2026 can be re-tested without that scaffolding. And OpenAI documents tool calling for this model on the Responses API only — tool calls issued through Chat Completions should move (model page).

The rates

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 for all four.

Input is half the old rate and output about 58% cheaper, which is deeper than the "50% lower" in OpenAI's launch post. Cached input fell to a quarter. Cache writes on the default 30-minute lifetime are $1.25 per million tokens, 1.25x input.

What one workload costs

Cost one out. A document pipeline: 10,000 invoice-extraction requests, each with a 2,500-token stable prompt (instructions and schema, cached after the first request), 1,500 fresh input tokens for the page image and text, and 1,500 output tokens including reasoning. Totals: 15M fresh input, 25M cached input, 15M output.

These are computed estimates on identical token counts, not measurements. Cache writes are excluded: the prompt is rewritten once per 30-minute window, and each 2,500-token write is under a cent. In every row output is most of the total, which is why the output rate matters more than the headline input cut.

Effort is the variable, not the price

The estimate above assumes both models burn the same tokens. They don't. Artificial Analysis (AA) measured GPT-6.1 Terra at about 6k output tokens per task at medium effort and 35k at max (AA report). A pipeline that moves from medium to max can multiply its output bill several times, and that is the single easiest way to lose the price cut.

The same report puts low and medium on AA's cost-efficiency frontier: at medium, an index score of 41 at $0.10 per task; at max, 48 at $0.40. Max buys a higher index score for four times the money.

What the benchmark charts claim

Every figure below is OpenAI's own, read from its launch charts as transcribed by Kingy AI, and none of it has been reproduced independently.

Higher is better except for the factual-error rows. The effort column matters as much as the score: on MMMU-Pro the new model reaches 84.6% at medium, already above the old model's max, and on OSWorld it reaches 66.7% at medium against 57.1% at max on the old model. The factual-error set is built from prompts users flagged, and the system card warns against reading those rates as production error rates.

Two HealthBench subsets went the other way: Professional rose from 57.7 to 62.7, while HealthBench fell from 57.0 to 55.1 and HealthBench Hard from 32.7 to 31.4.

Artificial Analysis's independent run agrees on direction and adds the number that matters for extraction work: its hallucination measure fell from 88% to 38% while accuracy moved from 47% to 49%, so far more of the new model's non-correct answers are abstentions rather than confident errors.

Three failure modes

1. Sampling parameters with reasoning on. With any effort other than none, OpenAI's migration guide says to drop temperature, top_p and top_logprobs; the Responses API rejects the request. Clients that hardcode a temperature value will get a 400 the moment they set an effort level. Fix: branch on effort, or pin effort to none for temperature-sensitive prompts and accept the weaker reasoning.

2. Tools on the wrong endpoint. Tool calling is documented for the Responses API only, so a Chat Completions function-calling loop is the second thing to audit. Fix: move the loop to Responses before pointing it at Terra.

3. The 272K line. Past 272,000 input tokens the whole request bills at 2x input and cache rates and 1.5x output. The 1.05M window is usable but not at the headline price, so a long-document path that silently crosses it pays double on input and cache and 1.5x on output. 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. Fix: measure input length at the top of the distribution per endpoint, not the average.

A fourth is not a bug but a budgeting trap: OpenAI's Batch and Flex processing take 50% off and Fast mode costs 2x, but those are OpenAI's offers. AIHubMix does not serve OpenAI's Batch API and its Terra price list shows standard rates only, so plan on standard pricing there.

Where the swap makes things worse

A cheaper token is not a cheaper task. On Terminal-Bench 4.0, Sol at medium scores 52.7% at an estimated $1.02 per task while Terra at max scores 50.3% at $1.84. On OSWorld 2.1 at max effort the per-task cost is $1.19 for Terra against $1.15 for Sol. On AA's index, Terra at max and Sol at medium both score 48, at $0.40 and $0.21. If a workload only passes at max effort, benchmark it against Sol at medium before migrating.

The system card adds a non-cost reason to test before switching agents: in a controlled test of respecting warnings, Terra at max worked around stated restrictions in 29.6% of rollouts, and 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 (system card).

A working request

AIHubMix's checks on 27 October found gpt-6.1-terra answering on Chat Completions, Responses and Anthropic Messages, at OpenAI's standard rates (model page). Reasoning effort is the reason to prefer Responses: all six levels were accepted there, 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), so don't trust it as a cost lever.

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)

Endpoint details are in the AIHubMix Responses API docs.

What can and cannot be verified

Measured, on 27 October: the three endpoints above answer, the six effort levels are accepted on Responses, moving from none to max raised reasoning tokens on the prompt used, and prices match OpenAI's standard list. Reproducible from the request above.

Vendor-reported, not reproduced here: every benchmark figure in this post comes from OpenAI's launch charts as transcribed by Kingy AI, or from the system card. The gains are consistent in direction with AA's independent run, but the numbers differ — AA ran Terminal-Bench 4.0 with mini-swe-agent on 66 tasks and got 35.4% for GPT-5.6 Terra and 55.6% for GPT-6.1 Terra, against OpenAI's 21.5% and 50.3%.

Not covered by any of it: your token counts. Set effort to medium, run your own task set, and compare cost per accepted result rather than cost per token.

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