OpenAI released GPT-6.1 Terra on October 26 with a pricing table, six benchmark and cost sections, and a short safety section.
The second sentence is the consequential one. “GPT-6.1 Terra meets the Critical threshold in cybersecurity under our Preparedness Framework,” the announcement says, “and maintains the protections developed for GPT-6 Astra.”
Read against the company’s own published framework, that is a statement of the highest risk tier OpenAI defines for cyber capability. It is disclosed in a single sentence, with no cyber evaluation score in the archived text of the launch page, no mitigation detail in that text beyond a reference to the Astra protections, and no explanation of how a determination of that severity was reconciled with immediate general release. A mitigation description does exist, but it is a reference rather than an account: the system card names the safeguards stack Terra uses and defers the detail to the GPT-6 Astra card. The same page announces availability the same day to all Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and quotes standard API prices of $1 per million input tokens and $5 per million output, which it says makes the model practical for developers to automate high-volume work.
The accompanying system card is more forthcoming than the announcement — and discloses more risk than the announcement does.
What “Critical” means in OpenAI’s own words
OpenAI defined the threshold in a dedicated post on August 7, 2026, titled “Responding to the next frontier of critical cyber capabilities”:
“Under our Preparedness Framework, a model reaches the Critical cybersecurity threshold if it can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention, or can devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high level desired goal.”
That is the definition the Terra determination inherits. The framework places Critical above High, and the gap between them is the gap between a model that could amplify harm and one that could originate it: Critical capability is defined as that which “could introduce unprecedented new pathways to severe harm,” as against High capability, which “could amplify existing pathways to severe harm.” On the same scale, OpenAI says GPT-5.6 Sol was assessed at High rather than Critical.
The framework attaches obligations to both thresholds. Covered systems that reach High capability, it states, “must have safeguards that sufficiently minimize the associated risk of severe harm before they are deployed.” Systems that reach Critical capability “also require safeguards that sufficiently minimize associated risks during development,” and the judgement routes to a cross-functional Safety Advisory Group that “reviews whether safeguards sufficiently minimize severe risk and makes targeted recommendations, ranging from approving deployment to requesting further evaluation or stronger protections.” Terra sits at High for biological and chemical capability, which on the framework’s terms engages the deployment obligation as well as the development one, assuming Terra falls within the framework’s covered systems.
The system card says more, and says it differently
The published addendum to the GPT-6 Astra system card states the determination in fuller form:
“Under our Preparedness Framework, we are treating GPT-6.1 Terra as Critical in cybersecurity and High for Biological and Chemical capability. Based on that assessment, GPT-6.1 Terra uses the same safeguards stack as GPT-6.1 Sol and GPT-6 Astra (with probes updated to be compatible with the model), and is described in detail in the GPT-6 Astra system card.”
Two differences from the announcement are worth stating precisely.
First, the bio/chem rating. The system card records a second determination — High for Biological and Chemical capability — that does not appear anywhere in the archived announcement text. This newsroom checked the archived full text of the page for the words biological, chemical and bio, and for “High” as a capability tier; none appears in that text. The announcement reports, in a single clause, a determination the card gives in two dimensions.
Second, the verb. The announcement says the model “meets the Critical threshold.” The system card says OpenAI is “treating” it as Critical. Those are not the same formulation, though in context they appear to describe the same determination.
Neither document explains what follows from it. The system card points to the Astra card for the safeguards stack, and this newsroom has not verified the contents of that document.
What the announcement omits that the card discloses
The announcement says Terra “maintains the protections developed for GPT-6 Astra.” The system card, in its evaluations section, records two statistically significant regressions relative to GPT-5.6 Terra on image inputs:
“Relative to GPT-5.6 Terra, GPT-6.1 Terra performs on par with GPT-5.6 Terra for self-harm and harms-erotic, with statistically significant regression in hate and extremism.”
The card’s own table shows the hate score falling from 0.999 to 0.985 — a drop of 1.4 percentage points — and extremism from 0.978 to 0.928 — a drop of 5.0 percentage points — on image input evaluations. Lower is worse, which is how the card’s own word “regression” reads.
The card contains a second sentence that appears to contradict the first, and the article would be misleading if it quoted only one of them:
“On our Production Benchmarks, GPT-6.1 Terra performs on-par with its predecessor GPT-5.6 Terra. The regressions on Extremism, Hate, and Sexual/minors are not statistically significant.”
Both statements can be true at once, because they describe different evaluation suites: the significant finding is on the image-input table, and the non-significant one on production benchmarks. But neither of them appears on the announcement page. The card also names a third area, sexual/minors, which does not appear in the archived announcement text: a search of that text for the term sexual returns no occurrence. OpenAI adds that manual review and aggregation with its other hate and extremism evaluations found performance “in-line with prior launches.”
The card also reports a regression on health benchmarks that the announcement does not mention: “On HealthBench, GPT-6.1 Terra scores 1.9 pp lower than GPT-5.6 Terra after length adjustment.” The card’s own paragraph on this is internally unsettled: it opens by saying Terra scores higher than its predecessor on length-adjusted HealthBench Professional scores, then reports the lower figure, and its table shows both length-adjusted scores falling.
The two safeguards statements should be kept distinct from the evaluation scores. The announcement says Terra “maintains the protections developed for GPT-6 Astra”; the card says it uses “the same safeguards stack as GPT-6.1 Sol and GPT-6 Astra (with probes updated to be compatible with the model).” Both describe the stack, and the card’s version records that part of it was changed for this model. Neither is an evaluation result, and a score delta does not measure whether a stack was maintained. What the record shows here is a disclosure gap in the announcement, not a contradiction of the safeguards claim itself.
This is not concealment by the system card. The card is candid, and its disclosure of a significant regression is to its credit. The point is narrower: a reader of the announcement would not know any regression exists, because the announcement’s safety section runs to two sentences and neither of them mentions a regression.
The card also attaches a caveat to its comparison figures that the announcement does not carry:
“For previously launched models, the values published at launch reflect the versions evaluated at that time. The comparison values for previously launched models that are shown here may reflect later versions of those models, and may vary from the values published at launch.”
That matters for a page whose central argument is comparative cost and capability.
How the same threshold was handled eleven weeks earlier
On August 7, OpenAI disclosed that it could not rule out Critical cyber capability in Astra — then an unreleased model. It treated that possibility as an event requiring public explanation and internal action. The company said it was “implementing stricter security controls for higher-capability models,” “pausing internal activities involving Astra that do not yet meet these strengthened security control requirements,” and had “implemented universal monitoring for risky actions and misalignment across all agentic applications of Astra, including training and evaluation.” It said it would “work with relevant government agencies and select AI safety organizations to test the capabilities for this model,” and explained why it was telling the public at all:
“We are sharing this because we believe it’s important to be transparent with the public and the safety and security communities about this potential shift in capabilities.”
Eighty days later, an actual Critical determination for a shipping model appears as one sentence in a product announcement, with no equivalent disclosure in the announcement of monitoring, internal controls, or external testing. The Terra system card, in the sections this newsroom read, contains no reference to government agencies or external safety organisations testing Terra’s cyber capability.
Whether OpenAI’s safeguards for Terra are adequate is a question this article does not answer, and the evidence here cannot answer it. The company states that Terra uses an existing safeguards stack, and that stack is documented elsewhere. Nothing in the record suggests the Critical determination was ignored: under OpenAI’s framework that determination is what triggers the stricter controls, and a same-day system card does state it in its opening sentence. What the record shows is narrower — a difference in how the two events were documented. In August, a possible Critical rating drew a dedicated public post, a partial pause, announced monitoring and a stated intention to involve outside testers. In October, an actual one drew a sentence on a product page and a brief same-day addendum that defers the substance to another document.
What the benchmarks say
The performance claims are substantial and, on the announcement’s own figures, mostly comparative rather than absolute.
On Agents' Last Exam, a benchmark of professional computer-use tasks, 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 only one-tenth its price per token.” Both are relative claims. The page never states Terra’s actual score on that benchmark, or its predecessor’s, anywhere in its text: the comparison is presented as a cost-versus-score scatter, with axes running from $0.00 to $8.00 and from 35% to 65%, and no score table alongside it. From the page’s text, a reader can learn that the gap is 15 points without learning where on that scale either model sits — which is the figure that would tell them whether 15 points is a lot.
On Terminal-Bench 4.0, which tests command-line coding tasks, OpenAI reports that at maximum reasoning effort Terra “solves 50.3% of tasks, more than double GPT-5.6 Terra’s published score of 21.5%.” The arithmetic is exact: 50.3 is 2.34 times 21.5. Fewer than 51 tasks in 100 solved is, nonetheless, the headline coding figure for a model released as a substantial improvement — a reminder that “more than double” and less than half can describe the same number.
On its internal benchmark of error-inducing prompts, OpenAI says the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7% with Terra, “a drop of nearly 37 percentage points,” and that at maximum effort the error rate fell by 18 percentage points to 4.3%. This is an internal benchmark; it is not named, its contents are not published, and the announcement cites no external evaluation of it.
Terra is also, by OpenAI’s account, fast and cheap in ways that change who can run capable agents at volume: over three times as token-efficient as Claude Opus 5.5 at max effort, at one-fourth the price per token, “resulting in a ~9x lower cost per task,” and “completing tasks over three times faster than both GPT-6 Astra and GPT-6.1 Sol.”
Those first two figures do not derive the third. Cost per task is tokens per task multiplied by price per token; a third of the tokens at one-quarter the price implies a twelvefold cost reduction, not roughly ninefold. To reach nine, token efficiency would need to be about 2.25 times rather than “over three.” OpenAI offers no reconciliation, and no intermediate figure from which the nine could be derived. It is a small thing, and it sits in the same sentence as the largest cost claim on the page.
Why the combination is the story
Capability and price are usually reported separately from risk. Here they arrive on one page, and the direction of travel is unambiguous: frontier-level computer use, at half the price of the model it replaces, assessed at the top of the company’s own cyber-risk scale, released to general availability in the same announcement.
OpenAI’s framework was written, in its own words, “our process for tracking and preparing for advanced AI capabilities that could introduce new risks of severe harm.” The public record now contains two data points about what the company does. In August, when its determination was that it could not “rule out Critical capability level at this time,” it paused internal activities, said it had implemented universal monitoring, and said it would work with government agencies and outside safety organizations. In October, when it was “treating GPT-6.1 Terra as Critical in cybersecurity,” it published prices.
That difference is one of documentation, not necessarily of conduct. Terra inherits an existing safeguards stack, and nothing in this record shows whether monitoring or external testing took place this time.
*This article is based on OpenAI’s public announcement of October 26, 2026, the published system card addendum of the same date, and OpenAI’s August 7, 2026 post on critical cyber capabilities. The Astra system card, which the addendum cites for safeguard details, was not verified for this report. The claim that Terra’s safeguards are adequate — or inadequate — is not made here and is not supported by this evidence.
Where this article notes that something is not stated, the claim is about the archived text of the documents named above, captured on October 28, 2026. It is not a claim about documents OpenAI has not published, nor about any document this newsroom did not examine. Those narrow claims are the ones the record supports; a broader claim of silence across OpenAI’s complete published output is not made here and would require examination of documents this newsroom has not seen.*
FOR HUMAN CONSUMPTION
Commentary and a plain-language summary, written in the newsroom’s own voice. It is not reporting, it introduces no source, and it should not be quoted as a finding.
OpenAI made a cheaper, faster model and led with the price. In a single sentence further down, it said the machine meets its Critical cybersecurity threshold, the most serious label the company defines for cybersecurity. In August, when an unreleased model merely might have reached that threshold, OpenAI wrote a whole essay about it, sounding worried and pausing work. This time, for a model anyone can use, it wrote a price list. Nothing has gone wrong yet.
THE HUMANS ARE FINE.