GPT-6.1 Terra Doubled Its Terminal-Bench Score and Halved Its Price — the Independent Run Finds It Fast, Cheap, and Inconsistent

OpenAI shipped GPT-6.1 Terra on October 26 at $1/$5 per million tokens, with its own terminal-line score more than doubling. An independent 101-task accounting run scores it 78.2% — 18th of 22 — with the second-lowest consistency in the field.

OpenAI released GPT-6.1 Terra on October 26, 2026, and the release reads like a clean win: the terminal-line benchmark more than doubled, from a published 21.5% to 50.3%, and the price fell by half to $1 per million input tokens and $5 per million output tokens. A mid-tier model that got meaningfully better and meaningfully cheaper is the kind of release that changes a line item on a bill.

On the same day, an independent benchmark of 22 models put Terra at 78.2% — tied for 18th — with the second-lowest consistency score in the entire field. 78.2% is a mid-tier grade. The consistency number is the one that should stop you before you route production work at it.

Hold three things while you read this: OpenAI’s benchmark doubling is a measurement of what the model can do once, on the vendor’s own harness. The consistency figure is a measurement of what it does every time, by somebody else. And they describe two different products, which is exactly where the review workload lives.

What OpenAI Actually Shipped

Terra is the middle rung of OpenAI’s line-up: cheaper and more capable than GPT-6 Luna, cheaper and less capable than GPT-6.1 Sol and GPT-6 Astra. It is available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and through the API as gpt-6.1-terra.

Price GPT-6.1 Terra
Input $1.00 / million tokens
Cached input $0.05 / million tokens
Output $5.00 / million tokens

OpenAI says those standard prices are 50% below GPT-5.6 Terra and GPT-6.1 Sol, and that Terra is around 10x the price of GPT-6 Luna. The arithmetic checks out against the announcement, and it is the one claim in the release that a reader can verify without running anything.

The emphasis in the announcement is unusual for a mid-tier model: detailed visual reasoning and precise computer use, with examples like interpreting technical drawings, filling out forms, editing video, and drafting proposals. The API documentation gives Terra a 1.05M context window.

Simon Willison’s read is that Terra occupies the same slot in OpenAI’s line-up that Sonnet occupies in Anthropic’s — “not the cheapest or the most expensive, making it harder to decide when to use it.” That is a better description of the product than the word “mid-tier.” A slot is easier to buy than a tier: the decision you are making is not “which model do I buy” but “which of my four jobs deserves the expensive one.”

The Vendor Numbers, With Their Conditions Attached

Everything below comes from OpenAI’s own announcement page. The coverage reproduces it: 9to5Mac quotes the error-rate figures verbatim, VGTimes carries the Terminal-Bench numbers, and Neowin repeats the mid-reasoning and cost-efficiency claims.

Vendor benchmark GPT-6.1 Terra Compared against Condition
Terminal-Bench 4.0 50.3% GPT-5.6 Terra’s published 21.5% max reasoning effort
Error-inducing prompts 16.7% error rate 53.6% for GPT-5.6 Terra low reasoning effort
Error-inducing prompts 4.3% error rate down 18 points; “nearly matching” Astra and Sol max reasoning effort
Agents’ Last Exam +15 points vs GPT-5.6 Terra at about one-third the cost medium reasoning effort
Terminal-Bench-Science 0.1 above GPT-6.1 Sol at lower cost within five points of Claude Opus 5.5 max reasoning effort; Opus cost is OpenAI’s estimate
MMMU-Pro above GPT-5.6 Terra at every setting “nearly matches” Astra and Sol OpenAI’s numbers

Four things to keep with that table.

The doubling compares against a published number, not a same-day rerun. OpenAI’s own wording is that Terra solves 50.3% of Terminal-Bench 4.0 tasks, “more than double GPT-5.6 Terra’s published score of 21.5%.” Nobody has published a run of both models on the same harness on the same day. Bench results are harness-and-scaffold dependent, and we measured a 29-point swing on SWE-bench Pro from the wrapper alone with nothing changed but the harness (TopClanker, August 10). A comparison to a number published under conditions you do not have is a comparison to a memory.

“9x lower cost per task than Claude Opus 5.5” is a vendor’s own arithmetic. The claim decomposes into two vendor facts — Terra is over 3x as token-efficient at max effort, and one-fourth the price per token — multiplied together, on OpenAI’s own Agents’ Last Exam harness. Neowin’s write-up is careful to present it as OpenAI’s figures rather than a third-party cost comparison, and VGTimes flagged the same numbers as not independently verified. They are not wrong; they are one company’s measurement of itself.

One of the two benchmark numbers has an independent counterpart, and it is higher. Artificial Analysis has run Terra on Terminal-Bench 4.0 and lists it at 55.6%, against OpenAI’s 50.3%. On MMMU-Pro the same source lists 84.7% against OpenAI’s 85.5%. Both numbers are tabulated by the BenchLM aggregator, which also lists Artificial Analysis’s AutomationBench (60.2%), SciCode (55.1%), Briefcase (1402) and GDPval-AA (1468) results for Terra — benchmarks where OpenAI publishes no comparable figure of its own. On the two benchmarks where an outside run exists, Terra lands within a few points of the vendor’s claim, one above and one slightly below. That is the most reassuring evidence in this post, and it is also the only outside evidence for the vendor’s headline number.

A 16.7% error rate on error-inducing prompts is still a 16.7% error rate. The internal error benchmark is the most interesting thing OpenAI published — a 37-point drop at low reasoning effort, and 18 points at max — but it measures error rate on prompts built to cause errors, on an internal set that nobody outside can run. It is a floor on hallucination, not a ceiling on reliability.

One more vendor line deserves to be quoted rather than paraphrased, because OpenAI put it in its own announcement: “For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.” The company that published the 50.3% also published the ceiling on the model.

The Independent Run: 101 Accounting Tasks

The DualEntry Accounting AI Benchmark tests models on 101 accounting tasks across eight areas — journal entries, accounts payable, accounts receivable, bank reconciliation, financial reporting, month-end close, transaction classification, and one more. Terra scored 78.2%, tied for 18th of 22 models.

Model Accuracy Passed all 3 runs Per task
Gemini 4 Argon 91.1% 86.1% —
GPT-6.1 Sol 81.2% 68.3% $0.09
GPT-6 Astra 81.2% 71.3% $0.49
GPT-6.1 Terra 78.2% 65.3% $0.042
Claude Haiku 5.5 76.2% — —
GPT-6 Sol 76.2% — —
GPT-6 Luna 66.3% lowest of the 22 —

The result was distributed the same day as a wire story through Bizwire (FinancialContent), so you will see it credited to several outlets. It is one study, run once.

Read the incentives before you read the number. DualEntry sells accounting automation software, and this is its own benchmark, published with a buyer-facing frame. That makes it a vendor’s test of a competitor’s model — closer to a useful independent check than to a neutral referee, and worth knowing before you treat 78.2% as a verdict on anything. It is also one workload in one domain. We are not going to claim Terra is a bad model because of 101 accounting tasks. We are going to claim that it is the only independent measurement of this model we found, and that it disagrees with the shape of the launch story.

Where Terra Is Genuinely Strong

  • Transaction classification: 100% (13 of 13).
  • Journal entries: 92.3% (12 of 13).
  • Bank reconciliation: 83.3% (10 of 12) — behind only Kimi K3.1 Pro (100%) and Muse Wonder (91.7%), and matching Grok 4.7 and Claude Fable 5.1.
  • Speed: 16.1 seconds per task, the fastest OpenAI model DualEntry has tested and third fastest overall, behind Claude Haiku 5.5 (11.4s) and Claude Sonnet 5.5 (12.1s) — roughly half the time of GPT-6.1 Sol (27.2s) and GPT-6 Astra (34.3s).
  • Cost: $0.042 per task, or $12.73 for a full three-run benchmark. Seventh cheapest of 22, roughly a tenth of GPT-6 Astra’s $0.49 and half of GPT-6.1 Sol’s $0.09.

One caution on the speed number. It is a per-task average on structured accounting work, and it is not a latency figure for a single prompt. Willison’s pelican render of Terra at max reasoning cost 11.3161 cents and took 4 minutes 11 seconds — the same model, on one request, at the reasoning setting most people will not use for bulk work and most agent loops should not either.

And Terra is not the best value per point. DeepSeek V4.1 Flash, GLM-5.3 Flash and GLM-5.3 all cost less per task and score higher; Mistral Large 4 costs about the same while scoring two points higher. Terra’s pitch is speed and price per task at mid-tier quality, not value.

The Profile Hides Inside the Average

Here is the most useful thing in the independent run, and it is not in the headline number.

Task difficulty Terra Rank in field
Easy 28 of 37 (75.7%) only GPT-6 Luna scored lower — roughly 21st of 22
Medium 31 of 36 (86.1%) only eight models beat it — about 9th of 22
Hard 20 of 28 (71.4%) only seven models beat it — about 8th of 22
Overall 78.2% tied 18th of 22

Terra is a top-ten model on medium and hard accounting work and close to the bottom on easy work. That is not a normal profile. Most models are better at hard tasks than easy ones, because the easy ones are the ones the benchmark defines as routine. Terra’s rank collapses on the routine band.

The benchmark’s write-up calls this the unusual weakness, and notes Terra “lost most of its points on simple cases, not hard ones.” On raw miss rate that framing is arguable — 9 of 37 easy tasks missed (24.3%) against 8 of 28 hard ones (28.6%) — but on rank it is exactly right, and rank is what the aggregate is made of. The 18th-place overall score is not “Terra is mediocre at accounting.” It is “Terra is excellent at the hard cases, good at the medium ones, and surprisingly weak at the ones everybody finds easy,” and the easy band is the largest single block of the test.

For an agent deployment, that is the worst possible shape. The hard cases are one-off. The routine cases are the ones you run over and over, and they are the ones Terra is least reliable on.

The Consistency Number Is the One to Plan Around

Terra passed 65.3% of tasks in all three benchmark runs — the second-lowest of the 22 models tested, ahead of only GPT-6 Luna. For comparison: GPT-6.1 Sol 68.3%, GPT-6 Astra 71.3%, Gemini 4 Argon 86.1%.

What that means operationally: for a task type the model has demonstrated it can do, it does it every time about two runs in three. If failures were roughly independent, a 65.3% all-three-runs rate implies a per-run success rate near 87% (0.653^(1/3)), which is about one run in eight failing on work the model has already shown it can do. The independence assumption is doing real work in that calculation and real workloads are not independent, so read it as a scale rather than a constant.

The benchmark’s own framing is the correct one: a low consistency score means the model is less likely to get the same work right every time, “which turns into review work for any agent that repeats a task.” That is the whole product decision. You can buy the cheapest model in the pack, and if it fails one run in eight, your review queue is one-eighth of your task volume — and a review queue that never shrinks is not a review queue, it is the workload.

Terra is also the weakest in the field where it is most expensive to be weak: OpenAI’s own release pitches Terra for reading incoming invoices and entering their details into accounting software, and the weak bands on the independent run are payables, receivables, and month-end close.

Where Terra Sits in the Lineup

Willison’s framing — the same slot as Sonnet — is the useful one, and it is reinforced by OpenAI’s own product man, Ted Sanders, on Hacker News: Terra is “probably better than Astra and Sol at technical communication, similar on vision and computer use (but much faster!), and a bit worse on most other things.” For API developers, he calls it “a nice midrange option if you want something cheaper than Sol but a lot smarter than Luna,” especially for vision, computer use, and communication. For subscribers, “not a bombshell release.”

The aggregator’s read is consistent and softer than either: BenchLM has Terra at 75.87/100, ranked #10 of 993 models, while disclosing that it covers 26 of 710 available benchmarks and that its own score is conservative as a result. An aggregate over partial coverage is an estimate, and an estimate over 26 of 710 benchmarks is an estimate of an estimate.

Verdict: What We Believe, in Order of Confidence

  1. The price cut is the durable story, and it is checkable. $1 per million input tokens and $5 per million output is 50% below the two models OpenAI prices Terra against, and the arithmetic holds without running anything. Whatever else is uncertain, the cost of high-volume structured work just dropped. That alone changes what you can afford to automate.
  2. The benchmark doubling is real but narrow. One benchmark, one vendor harness, max reasoning effort, compared against a predecessor’s published number rather than a same-day rerun — with the one independent counterpart we found (Artificial Analysis at 55.6%) landing above the vendor’s claim. Expect the direction to hold; do not expect the magnitude to survive a re-run under different conditions.
  3. The independent evidence does not support “Terra is a leader.” 78.2% and 18th of 22 on the only third-party measurement we found is a mid-tier score. It is also the fastest OpenAI model that benchmark has tested, at a tenth of Astra’s per-task cost — which is a much more interesting result than the rank.
  4. Plan around the profile, not the average. Terra is top-ten on hard work, near-last on routine work, and fails roughly one run in eight on tasks it has demonstrated it can do. If you put it on a repeatable workflow, you are buying throughput and paying in review. If you put it on one-off hard cases, where its accuracy is genuinely strong, it looks like the cheap tier it is.

What would change our mind: a consistency measurement on a second, non-accounting workload; an independent replication of the Agents’ Last Exam and Terminal-Bench figures under published harness conditions; or a vendor-neutral write-up of Terra’s reliability profile. We would also want to see Terra run on a task family with hundreds of instances rather than 13, because on this evidence a 13-task cell cannot tell you much about a repeat workflow — including the part of the benchmark where Terra is worst.

Sources