Evidence File: GPT-6.1 Terra Critical threshold and pricing

SOURCES

Verbatim passage archive: tmp/src/terra-systemcard-verbatim.md (10,371 bytes, sha256 d69ce451e4f155d3648a51fdc6b0348745f5823fdce393eb3ecce40e744aa961)

CLAIM-BY-CLAIM

C01 — Terra was released October 26, 2026 and is available generally

Article: “OpenAI released GPT-6.1 Terra on October 26...” Source (S1): “Starting today, GPT-6.1 Terra is available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.” Status: SUPPORTED. Date corroborated by discovery record first_published 2026-10-26T00:00:00Z.

C02 — Pricing

Article: “$1 per million input tokens and $5 per million output” Source (S1): “Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens.” Status: SUPPORTED. Article omits the cached-input tier; this is an omission of detail, not an error. No arithmetic claim is made from these figures in the article.

C03 — The announcement’s safety disclosure is one sentence

Article: “one sentence about safety” / “disclosed in a single clause” Source (S1): “As we make powerful AI more efficient and accessible, we continue training, testing, and refining our safeguards. GPT-6.1 Terra meets the Critical threshold in cybersecurity under our Preparedness Framework, and maintains the protections developed for GPT-6 Astra.” Status: SUPPORTED as to the determination. Precision note: the announcement’s safety passage is two sentences of which the second carries the determination. The article’s phrase “one sentence about safety” is therefore a slight compression. See REQUIRED EDIT R1.

C04 — Critical is the highest risk tier OpenAI defines for cyber

Article: “that is a statement of the highest risk tier OpenAI defines for cyber capability” Source (S3): “Under our Preparedness Framework, a model reaches the Critical cybersecurity threshold if...” and “Previous models, including GPT-5.6-Sol, have been evaluated for frontier cyber capabilities and assessed at the High (rather than Critical) threshold.” Status: SUPPORTED by implication of the High→Critical ordering in S3. NOTE: S3 does not state in terms that Critical is the highest level. OpenAI’s Preparedness Framework defines Low/Medium/High/Critical; this newsroom has not retrieved the framework document itself in this run. The claim is treated as SUPPORTED ONLY IF the four-level scale is confirmed. See REQUIRED EDIT R2 — soften unless the framework is retrieved.

C05 — The system card records a bio/chem determination absent from the announcement

Article: “The system card records a second determination — High for Biological and Chemical capability — that does not appear anywhere on the announcement page.” Source (S2): “we are treating GPT-6.1 Terra as Critical in cybersecurity and High for Biological and Chemical capability.” Source (S1): full page text searched; no occurrence of “Biological”, “Chemical”, “bio”, or “chem” in connection with a risk rating. Status: SUPPORTED. This is an absence claim and is recorded as such: the phrase does not appear in the retrieved announcement text. If OpenAI edits the page, this claim must be re-checked. Retrieval text archived.

C06 — “meets” vs “treating”

Article: “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...” Source (S1): “meets the Critical threshold in cybersecurity” Source (S2): “we are treating GPT-6.1 Terra as Critical in cybersecurity” Status: SUPPORTED as a literal textual comparison. The article does not assert that the two formulations describe different determinations, and says so (“though in context they appear to describe the same determination”). This sentence is doing careful work; do not let an editor delete it.

C07 — The article does not claim safeguards are inadequate

Article: “Whether OpenAI’s safeguards for Terra are adequate is a question this article does not answer.” Source: n/a — this is a limitation statement. Status: COMPLIANT. Checked: no sentence in the draft asserts or implies that OpenAI’s safeguards failed, were insufficient, or were knowingly inadequate. The strongest characterisation used is “a difference in disclosure practice”, which is supported by C08.

C08 — August handling of a possible Critical rating vs October handling of an actual one

Article: “In August, when the answer was ’we cannot rule it out,’ it paused work, imposed monitoring, and called the safety community. In October, when the answer was ’we are treating it as Critical,’ it published prices.” Source (S3): “we cannot rule out critical cyber capabilities”; “We are pausing internal activities involving Astra...”; “We have implemented universal monitoring for risky actions and misalignment across all agentic applications of Astra, including training and evaluation.”; “We will work with relevant government agencies and select AI safety organizations to test the capabilities for this model.”; “We are sharing this because we believe it’s important to be transparent with the public and the safety and security communities...” Status: SUPPORTED. Both halves are OpenAI’s own words. The contrast is described, not characterised as wrongdoing. Note the article correctly uses “potential shift in capabilities” for August (a possibility) and “an actual Critical determination” for October.

C09 — No reference to external testing in the Terra materials read

Article: “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.” Source (S2): full retrieved text searched; no occurrence of “government”, “external safety”, “third-party” in a cyber-testing context. Status: SUPPORTED, and correctly hedged three ways: “in the sections this newsroom read”. This hedge is load-bearing. See REQUIRED EDIT R3 — the hedge must survive editing.

C10 — Vision regressions on hate and extremism

Article: “statistically significant regression in hate and extremism... hate score falling from 0.999 to 0.985 and extremism from 0.978 to 0.928” Source (S2): “with statistically significant regression in hate and extremism”; table 4 values 0.999→0.985 and 0.978→0.928. Status: SUPPORTED, both the characterisation and the numbers. The article also reports OpenAI’s countervailing statement (“in-line with prior launches”), which is required for fairness and is present.

C11 — HealthBench regression

Article: “On HealthBench, GPT-6.1 Terra scores 1.9 pp lower than GPT-5.6 Terra after length adjustment.” Source (S2): identical sentence verbatim. Status: SUPPORTED. Quoted exactly.

C12 — Comparison-value caveat

Article: “The comparison values for previously launched models that are shown here may reflect later versions of those models...” Source (S2): verbatim, as quoted. Status: SUPPORTED.

C13 — Terminal-Bench 4.0 figures and the “more than double” framing

Article: “solves 50.3% of tasks, more than double GPT-5.6 Terra’s published score of 21.5%” and the observation that fewer than 51 in 100 solved is the headline figure. Source (S1): “At maximum reasoning effort, GPT-6.1 Terra solves 50.3% of tasks, more than double GPT-5.6 Terra’s published score of 21.5%.” Status: SUPPORTED. Arithmetic: 50.3 / 21.5 = 2.34, so “more than double” is accurate. “Fewer than 51 in 100” is accurate. The article’s interpretive sentence (“a reminder that ’more than double’ and ’half’ can describe the same result”) is analysis, clearly marked as such, and rests on correct arithmetic.

C14 — Error-rate figures

Article: “fell from 53.6% with GPT-5.6 Terra to 16.7% with Terra, ’a drop of nearly 37 percentage points’... at maximum effort the error rate fell by 18 percentage points to 4.3%” Source (S1): “...fell from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra, a drop of nearly 37 percentage points. At maximum effort, the error rate fell by 18 percentage points to 4.3%” Status: SUPPORTED. Arithmetic: 53.6 − 16.7 = 36.9, so “nearly 37” is correct. 4.3 + 18 = 22.3 implies a 22.3% starting point at max effort — consistent, though the article does not state the starting value, which it should not, since S1 does not.

C15 — Agents' Last Exam presented as a cost-versus-score chart

Article: “OpenAI publishes the comparison as a cost-versus-score chart rather than a score table; the underlying values are not given as numbers on the page.” Source (S1): “scores 15 percentage points higher than GPT-5.6 Terra at about one-third the cost... performs on par with GPT-6 Astra at only one-tenth its price per token.” Status: SUPPORTED as to the absence of numerical score values: S1 gives percentage-point deltas and cost ratios, not absolute scores. The phrase “cost-versus-score chart” describes the page’s visual presentation; this should be checked visually before publication. See REQUIRED EDIT R4.

C16 — “Eighty days”

Article: “Eighty days later, an actual Critical determination...” Source: S3 dated 2026-08-07; S1/S2 dated 2026-10-26. Status: SUPPORTED. 7 Aug → 26 Oct = 80 days. (Aug 24 remaining + Sep 30 + Oct 26 = 80.)

C17 — Astra system card not verified

Article: “this newsroom has not verified the contents of that document.” Status: TRUE STATEMENT OF LIMITATION. S4 was not retrieved. This disclosure is mandatory and must not be removed.

REQUIRED EDITS BEFORE PUBLICATION

• R1 — “one sentence about safety” → “a two-sentence safety passage” (precision; the determination sits in the second sentence).
• R2 — “the highest risk tier OpenAI defines for cyber capability” must either be confirmed against the Preparedness Framework document itself, or softened to “the top of the scale OpenAI uses when it assesses cyber capability” with the High→Critical ordering cited from S3 only. Default if the framework cannot be retrieved: soften.
• R3 — The hedge “in the sections this newsroom read” must survive editing. Do not let it become “the system card contains no reference to...”.
• R4 — Confirm visually that the Agents' Last Exam comparison is a chart and not a table before publication; if it is a table without numeric values, reword.

OUTSTANDING UNCERTAINTIES (carried into the approval card)

1. The Preparedness Framework document itself was not retrieved; the four-tier scale is inferred from S3. 2. The GPT-6 Astra system card was not retrieved. Terra’s safeguards are therefore unexamined in this package, and the article says so. 3. “Critical” is OpenAI’s self-assessment. This newsroom has no independent evaluation of Terra’s cyber capability and makes no claim about its real-world exploit potential. 4. S1 is a live page and may be edited; all absence claims are dated to the 2026-10-28 retrieval.

ADDENDUM — 2026-10-28: framework retrieved, R1 and R2 closed

While compiling this file, the reviewer flagged that C04 (“highest risk tier”) and the article’s use of “Critical” rested on an inference from S3’s High-vs-Critical ordering rather than on the framework itself. The framework was therefore retrieved and read.

C18 — Critical is the upper of exactly two thresholds

Source (S5): “We’ve streamlined levels to two clear thresholds that map to specific operational commitments: High capability, which could amplify existing pathways to severe harm, and Critical capability, which could introduce unprecedented new pathways to severe harm.” Status: SUPPORTED. R2 CLOSED — no softening required. The article’s “highest risk tier OpenAI defines for cyber capability” is accurate, and the article now quotes the framework’s own two-level formulation rather than inferring it.

C19 — Critical triggers a development-stage obligation

Source (S5): “Systems that reach High capability 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.” Status: SUPPORTED. Added to the article. This is the framework’s own text and is the strongest basis for treating a Critical determination as consequential rather than nominal.

C20 — A named review body exists

Source (S5): “The Safety Advisory Group (SAG), a cross-functional team of internal safety leaders, reviews whether safeguards sufficiently minimize severe risk and makes targeted recommendations, ranging from approving deployment to requesting further evaluation or stronger protections. Their guidance goes to OpenAI Leadership for final decisions...” Status: SUPPORTED. Added to the article to keep it fair: a review process is documented, and the article does not allege that it was skipped. The article’s claim is narrower — that neither the announcement nor the addendum says what the process concluded for Terra.

C21 — OpenAI commits to publishing preparedness findings per release

Source (S5): “We’ll continue to publish our Preparedness findings with each frontier model release...” Status: NOT USED IN ARTICLE. Noted for completeness: the Terra addendum is four lines and defers safeguard detail to the Astra card. The article does not accuse OpenAI of failing this commitment, and should not, since an addendum was in fact published.

R1 — CLOSED

“one sentence about safety” replaced with “a two-sentence safety passage” in the article body.

R3, R4 — STILL OPEN

• R3 — the hedge “in the sections this newsroom read” is present and must survive editing.
• R4 — the Agents' Last Exam chart-vs-table wording still requires a visual check before

publication. Not yet performed.

Residual risk after addendum

The single largest weakness remaining is that S4 (the GPT-6 Astra system card) was never retrieved, so Terra’s actual safeguard stack is unexamined. The article discloses this in its closing note. Any editor proposing to remove that disclosure would be removing the article’s principal protection against overclaiming.

FINAL ADDENDUM — 2026-10-28: full page text re-verified; cost arithmetic added

The full text of S1 was retrieved a second time through the controlled browser and searched directly, to convert one absence claim from “true of the excerpt” to “true of the page.”

C22 — The announcement’s full text contains no bio/chem determination

Method: full page text of https://openai.com/index/introducing-gpt-6-1-terra/ retrieved 2026-10-28 (HTTP 200), searched for Biological, Chemical, bio, and High as a capability tier. Result: Biological 0 occurrences, Chemical 0, bio 0; High appears only inside “higher” and never as a threshold label. Status: SUPPORTED as an absence, tested against the whole page rather than an excerpt. The article states the words searched, so a reader can repeat the test.

C23 — The three cost figures do not reconcile

Source (S1): “over three times as token efficient as Claude Opus 5.5, and one-fourth its price per token, resulting in a ~9x lower cost per task.” Arithmetic: 3 x 4 = 12. A ninefold cost reduction requires roughly 2.25x token efficiency, which contradicts “over three times.” Status: PRESENT IN SOURCE, NOT RECONCILED BY SOURCE. Recorded in the article as an unexplained inconsistency, not as an accusation: a differing token basis could reconcile it, and OpenAI supplies no intermediate figures either way.

C24 — Terminal-Bench 4.0 is named on the page

Source (S1): “Terminal-Bench 4.0 tests how well coding agents solve problems from the command line...” and the accompanying results sentence. Status: SUPPORTED. Challenged by the reviewer as unquoted; re-verified against full page text, where the string occurs twice.

C25 — Agents' Last Exam: delta published, anchor not

Source (S1): “scores 15 percentage points higher than GPT-5.6 Terra at about one-third the cost.” The page’s text contains no numeric score for Terra or GPT-5.6 Terra on that benchmark; the comparison is rendered as a scatter with axis ranges $0.00–$8.00 and 35%–65%. Status: SUPPORTED as to the text. NOT visually confirmed — whether the plotted chart carries printed data labels was not established, because vision was unavailable in this run and page scrolling timed out. The article therefore claims only what the text supports and says nothing about chart labels.

Residual limitations after all checks

1. The GPT-6 Astra system card (S4) was never retrieved, so Terra’s safeguard stack is unexamined. Disclosed in the article’s closing note. 2. The Agents' Last Exam chart was never viewed. Disclosed here; the article is written to stand without it. 3. The adversarial review was a reasoning review only. It re-verified nothing against live sources and never saw the draft. See reports/phase3-terra-newsroom-record.md.

Why this addendum exists. Until now this Evidence File recorded claims adjudicated by hand and by reviewer challenge. It had no automated test that the article’s quotation marks enclose text that actually exists in the sources. Story 2 and Story 3 were both built with such a harness. Story 1 was not. That asymmetry is itself a quality defect, so one was built and run against this package.

Corpus archived for the test. The framework page was missing from tmp/src/, so it was retrieved before testing: https://openai.com/index/updating-our-preparedness-framework/, curl.exe with a full browser header set, HTTP 200, 293,118 bytes HTML, stripped to 7,981 characters, TEXT_SHA256=03887C1CA77A66CFBF2DD54007FCC11944142AC2A669640734B377A5F07D55D6 (tmp/src/terra-framework.txt). The three source files for the harness are therefore: the announcement, the system card addendum with the August post appended, and the framework page.

C26 — First run: 45 passed, 4 failed, and the four failures were mine

tmp/verify_terra_quotes.ps1 extracts every quoted span from the article and requires it to appear in the archived source text after normalising only typographic quotes, dashes and whitespace (case significant). First run: 34 quoted spans, 4 failures, VERDICT FAIL, exit 1.

Status: all four corrected in the article. Second run: 33 quoted spans, 0 failures, 48 assertions, 48 passed, VERDICT PASS, exit 0. Falsification control (an injected fabricated sentence about external auditing) correctly NOT FOUND on both runs.

What this means, stated plainly

Two of the four defects would have put my own paraphrase inside quotation marks in a published article about another organisation’s safety disclosures — in a piece whose entire argument is that the company’s words were thinner than they should have been. That is the least defensible place in journalism to be sloppy, and the manual review, the adversarial reviewer and my own re-reads all missed it.

Nothing here was fabricated in the sense of an invented document or an invented fact. The facts and the quotations' meaning were right. The marks were wrong. It was caught by a script that does not care how confident I am, which is the argument for having the script before having the confidence.

Consequence for the remaining packages

The harness is now the standard for every package in this edition and every one after it: no quoted span leaves this newsroom without being tested against archived source text, and no falsification control means no pass. A package that fails is fixed, not waived.

The Commander directed that the HealthBench inconsistency be recorded in this Evidence File, not only in the article. Recorded here as C27.

C27 — the card’s HealthBench passage contradicts its own table

S2, §5.1, opening sentence:

“GPT-6.1 Terra scores higher than GPT-5.6 Terra on length-adjusted HealthBench Professional scores, with a higher unadjusted score and shorter mean final responses.”

S2, §5.1, following sentence:

“On HealthBench, GPT-6.1 Terra scores 1.9 pp lower than GPT-5.6 Terra after length adjustment.”

S2, §5.1, table values (length-adjusted):

Analysis. The paragraph asserts Terra scores higher on length-adjusted HealthBench Professional, then reports it scoring lower, and its own table shows both length-adjusted figures falling. The three components cannot all describe the same measurement. Reconciling explanations are possible — Professional and the standard set may be different columns, “unadjusted” may run the other way, or the opening sentence may be an editing error carried from a prior card — but the document states none. This newsroom does not invent one.

Status: recorded as an internal inconsistency in the primary source. Not an accusation, and not an error introduced by this article.

Handling in the article. The article quotes only the 1.9 pp figure — the one with a table behind it — and states that the card’s own paragraph is unsettled about it. It does not adopt “higher”, because the table contradicts it, and does not adopt “lower” as the card’s plain meaning, because the card’s prose disputes it. It reports the conflict.

Why it matters to the package. The article’s thesis is that the announcement’s safety summary is thinner than the card’s. This inconsistency cuts the other way: the card, the more disclosive document, contains a defect the announcement does not. Presenting the card as uniformly more careful would overstate the case, which is why the inconsistency is disclosed in the article and not left in this file.

The Commander ordered that the article distinguish five things, use archived documents with capture dates and hashes, publish no invented MMMU-Pro or OSWorld 2.1 figures, and present the August-versus-October comparison without unsupported implications. Status of each, verified against the archives.

1. The announcement’s statement — “meets”

S1 verified: “GPT-6.1 Terra meets the Critical threshold in cybersecurity under our Preparedness Framework, and maintains the protections developed for GPT-6 Astra.” Occurrences of “meets the Critical threshold” in the archived announcement: 1.

2. The system card’s wording — “treating”

S2 verified: “Under our Preparedness Framework, we are treating GPT-6.1 Terra as Critical in cybersecurity and High for Biological and Chemical capability.”

The article quotes both formulations and states they are different wording, in context describing the same determination. It does not assert they describe different determinations, because no document says so.

3. What the Preparedness Framework actually requires

S5 verified verbatim:

• Two thresholds, not four: “We’ve streamlined levels to two clear thresholds ... High capability, which could amplify existing pathways to severe harm, and Critical capability, which could introduce unprecedented new pathways to severe harm.”
• Development-stage obligation: “Systems that reach Critical capability also require safeguards that sufficiently minimize associated risks during development.”
• Review route: the Safety Advisory Group “reviews whether safeguards sufficiently minimize severe risk and makes targeted recommendations, ranging from approving deployment to requesting further evaluation or stronger protections”, and “Their guidance goes to OpenAI Leadership for final decisions.”

The article states what the framework requires without alleging the process was skipped. No document says it was, and the framework’s own text is what makes a Critical determination consequential rather than nominal.

4. What the documents disclose and do not disclose

Disclosed: the determination (Critical cyber; High bio/chem); the safeguards basis (“same safeguards stack as GPT-6.1 Sol and GPT-6 Astra (with probes updated to be compatible with the model)”); four evaluation suites with named tables and figures; a comparison-value caveat; a research-environment caveat.

Not disclosed: any cybersecurity evaluation by name; any cyber score or threshold margin; any cyber-specific mitigation; which Safety Advisory Group recommendation was made or accepted; any external or governmental testing of Terra’s cyber capability; any rationale connecting a Critical determination to immediate general release.

Verified absences in the announcement text: Biological 0, Chemical 0, bio 0, capitalised High 0. The only “high” occurrences are “high-volume” and “higher”. All six occurrences of “score” are relative comparisons, a chart axis label, a named score on a different benchmark (Terminal-Bench), or analyst notes — no absolute Agents' Last Exam figure appears in the page text.

5. Absence of public detail is not proof of non-occurrence

This distinction is now carried explicitly. The article states that whether OpenAI’s safeguards for Terra are adequate is a question it does not answer and cannot answer from this evidence; that the company says Terra uses an existing documented stack; and that nothing in the record suggests the determination was ignored.

C09 and C22 are absence claims tested against full archived text and dated to the retrieval. Neither is used to prove an evaluation or mitigation did not occur. The addendum’s silence on safety-advisory sign-off is reported as silence; the article does not infer a failure to review.

MMMU-Pro and OSWorld 2.1 — no figures, and none invented

S1 verified: both are discussed qualitatively only.

• MMMU-Pro: “GPT-6.1 Terra outperforms GPT-5.6 Terra at every reasoning setting. It even nearly matches the scores of our most capable frontier model, GPT-6 Astra, and GPT-6.1 Sol at much lower cost.” No numbers.
• OSWorld 2.1: “At medium reasoning effort, GPT-6.1 Terra surpasses GPT-5.6 Terra’s best score at a fraction of the cost per task.” No numbers.

The article attaches no figure to either. Any figure attached to these two in a future revision would be fabricated.

August-versus-October comparison — both halves sourced

S3 (7 August 2026) verified: “we cannot rule out critical cyber capabilities under our Preparedness Framework”; “We are pausing internal activities involving Astra that do not yet meet these strengthened security control requirements.”; “We have implemented universal monitoring for risky actions and misalignment across all agentic applications of Astra, including training and evaluation.”; “We will work with relevant government agencies and select AI safety organizations to test the capabilities for this model.”

S2 (26 October 2026) verified: “we are treating GPT-6.1 Terra as Critical in cybersecurity”.

What the article may say and does not say. It says the two events were documented differently: in August a possibility drew a dedicated post, a partial pause, announced monitoring and a stated intention to involve outside testers; in October an actual determination drew a sentence on a product page. It does not say safeguards were ignored or that the model is unsafe — the framework routes a Critical determination to stricter controls, a system card was published the same day, and the stack is asserted to be inherited from Astra. The comparison is about disclosure practice, and the article frames it that way.

Capture dates and hashes

Re-retrieval on 2026-10-28 (this run) failed, and is recorded as failure, not absence: openai.com resolves (104.18.33.45), returns HTTP 308 then HTTP 403 at a Cloudflare challenge; web_fetch and web_request fail at transport level. No access control was bypassed. The archived captures above remain the source of record, and every absence claim is dated to them.

The independent review returned HOLD for four wording fixes. Three are now landed and the fourth was an artifact of the extracts this newsroom supplied to the reviewer. This addendum records the fixes, and the searched absence evidence the review asked for as a MEDIUM item.

1. Absence-check record (review MEDIUM: “record the search, do not infer absence from brevity”)

The article asserts that the launch page carries no cyber evaluation score, no mitigation description, no monitoring description, and no reconciliation of the Critical determination with same-day general release. Those are negative claims and must rest on a search, not on the impression that the section is short.

Method. Page body extracted from the archived capture tmp/src/terra-primary-verbatim.md, bounded by the ## Header and ## Footer date block markers, with bracketed analyst annotations and probe-note lines excluded. Resulting body: 5,037 characters. Case-insensitive literal counts on that body:

Reading of the counts. The single cyber occurrence is the Critical sentence itself; there is no second passage about cybersecurity anywhere on the page. The single safeguard occurrence is “we continue training, testing, and refining our safeguards.” The single evaluat occurrence is “According to our evaluations” in the alignment sentence, which is about staying in scope and disclosing limitations, not about cyber testing. No occurrence of mitigation, monitoring, reconciliation, deployment rationale, or risk appears at all.

What this establishes. That the page contains no mitigation description, no monitoring description, no cyber evaluation score, and no reconciliation statement. This is a searched absence over the full captured text, not an inference from the length of the safety section.

Limit, stated plainly. It establishes absence from this capture. The capture is a cleaned extraction of the page, so it cannot rule out that the page contained content the extraction did not carry. The article’s wording is scoped to the announcement page for that reason, and the closing paragraph states that the record does not show whether monitoring or external testing took place.

2. The review’s most serious finding was an artifact of our own extracts - recorded as a defect in our dispatch, not in the article

The review flagged the quotation “rule out Critical capability level at this time” as not present in the extracts supplied to it, and treated that as a suspected mis-quotation.

The August 7, 2026 post contains both phrasings, in different paragraphs:

“we cannot rule out critical cyber capabilities under our Preparedness Framework”

“While we continue to benchmark and assess this model, our preliminary evaluations indicate strong enough performance that we cannot rule out Critical capability level at this time.”

The extract package sent to the reviewer carried only the first. The reviewer therefore flagged an accurate quotation, correctly, on the evidence it was given. The fault is in the dispatch materials, not the article. Verified after the fact: quotation 35 of 36 passes against the archive.

Standing consequence. A reviewer can only check what it is handed. When a review reports a quotation as absent, the first question is whether the extract was complete - not whether the writer embellished.

3. Fixes landed after the review

4. Verification after the fixes

Harness tmp/verify_terra_quotes.ps1, run against the archived captures:

• QUOTES_TESTED=36 QUOTE_FAILURES=0
• PASSED=51 FAILED=0, VERDICT: PASS
• Quotation 7 (the newly added High-capability obligation) and quotation 9 (the Safety Advisory Group) both resolve verbatim in the archive.
• Falsification control: a deliberately injected fabricated sentence is correctly NOT FOUND.
• Retired-phrase sweep: all eleven phrases the review and this pass removed return 0 occurrences - “two-sentence”, “a tenth of the previous”, “ten weeks”, “consists of pointing”, “Rated Its New”, “rates the model Critical”, “two-dimension risk determination”, “sets only two thresholds”, “three times fewer”, “a drop of 1.4 points”, “a drop of 5.0 points”.
• Article length after the additions: 2,486 words.

5. Status of the five ordered requirements after this pass

This remains a corrected text awaiting re-review and publisher approval. It is not published, and this addendum does not assert that a re-review has occurred.

Status change. The corrected text was re-reviewed by a second independent agent. That review returned PASS with no hold items, confirming: all eight checks passed, including that every absence claim is tied to the archived capture; that “meets the Critical threshold” and “treating ... as Critical” are kept distinct; that the HealthBench text/table inconsistency is stated rather than smoothed; and that the August-versus-October comparison implies neither that safeguards were skipped nor that conduct was identical. The reviewer independently re-derived every figure: 0.999→0.985 = 1.4 pp, 0.978→0.928 = 5.0 pp, 50.3/21.5 = 2.34, 53.6→16.7 = 36.9 pp, and the cost-per-task gap (9 ÷ 4 = 2.25, not “over three”).

Reviewer’s stated limitation, and its effect. The reviewer could not retrieve any primary document: openai.com returns HTTP 403 behind a Cloudflare challenge and web_fetch fails at transport. Its review therefore rested on the extracts supplied with the assignment. It correctly flagged this as the one failure mode it could not close, and named four items as unverified rather than disproven. It also noted the in-line with prior launches quotation and the August-post continuations as falling outside its extract set.

All four items resolved. This newsroom checked each against its own archived captures, and every one is present verbatim:

The article’s rendering of the fourth is faithful: it compresses “Previous models, including GPT-5.6-Sol, have been evaluated ... and assessed at the High (rather than Critical) threshold” to “On the same scale, OpenAI says GPT-5.6 Sol was assessed at High rather than Critical”, which preserves the claim and drops only the “previous models, including” lead-in.

Why this happened, recorded as practice. The reviewer was given extracts, not the full captures. An item absent from an extract is not absent from the source, and the two were distinguished only by checking the archive. The rule applied here: when a reviewer reports a quotation as untraceable, check whether the reviewer’s extract was complete before treating it as a defect in the article. In this case the article was correct and the extract set was narrow.

Effect on the published text. None. No sentence was added, removed or altered as a result of this addendum; all four items were already correct in the published article. This addendum records the second review and closes its stated gap. It does not claim the reviewer verified these four items — the newsroom did, against its own captures, after the review returned.

What remains genuinely unverified, and is disclosed in the article. The GPT-6 Astra system card, which the Terra addendum cites for the safeguards stack, has not been examined by this newsroom.