OpenAI launched GPT-6.1 Terra on Monday, October 26: standard API prices 50% below GPT-5.6 Terra and GPT-6.1 Sol ($1 / $0.05 / $5 per million tokens), with the vendor itself reporting the model at Critical — the higher of its own framework's two cybersecurity thresholds

2 sources

OpenAI introduced GPT-6.1 Terra on Monday, October 26, 2026 (the OpenAI news RSS pubDate for the post reads 'Mon, 26 Oct 2026 00:00:00 GMT'; the fetched openai.com/news listing shows the same card dated October 26, 2026). Per the official post, it is 'a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing', and the standard API prices it reports are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol: $1 per million input tokens, $0.05 per million cached input tokens, $5 per million output tokens. The vendor-reported benchmark set: on Agents' Last Exam at medium reasoning effort, 15 percentage points higher than GPT-5.6 Terra at about one-third the cost, and on par with GPT-6 Astra at one-tenth its price per token; at maximum effort, over three times as token efficient as Claude Opus 5.5 at one-fourth its price per token — OpenAI's own arithmetic 'a ~9x lower cost per task' — and over three times faster than both GPT-6 Astra and GPT-6.1 Sol. Error rate on OpenAI's internal benchmark of error-inducing prompts: 53.6% for GPT-5.6 Terra at low reasoning effort fell to 16.7% for GPT-6.1 Terra — the post's rendering is 'a drop of nearly 37 percentage points' — and at maximum effort it fell 18 points to 4.3%. Terminal-Bench 4.0 at maximum reasoning effort: 50.3% of tasks, 'more than double GPT-5.6 Terra's published score of 21.5%' — with the post's own boundary 'For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.' Terminal-Bench-Science 0.1 at maximum reasoning effort: surpasses GPT-6.1 Sol at lower cost, within five points of Opus 5.5 at less than one-fifth the estimated cost. Help Net Security (Anamarija Pogorelec, published_time 2026-10-27T04:33:34+00:00, opened via raw HTML) independently renders the same numbers and adds the outlet's frame for the safety claim: 'Critical is the higher of two cybersecurity thresholds in the version of the Preparedness Framework dated April 2025.' and 'On September 1, 2026, OpenAI designated Astra as its first model at this level.' — a designation history that makes Terra's Critical rating a self-designation under OpenAI's own framework, not an external audit. The outlet also carries the official workload example — 'The company's own example of high-volume work for Terra is reading incoming invoices and entering the details into accounting software.' — and the behavior deltas: 'Terra is more likely than GPT-5.6 Terra to stay within scope and to disclose limitations and blockers honestly.' Availability per the post: 'Starting today, GPT-6.1 Terra is available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Developers can also access it through the OpenAI API as gpt-6.1-terra.' Every benchmark figure in this card cross-checks between the official post's reader capture and Help Net Security's raw-HTML rendering — values agree, with two surface differences: the outlet spells out '50 percent' and '18 percentage points', and the 'nearly 37 percentage points' rendering has no outlet counterpart.

Why now

The launch landed Monday and the pricing is live in the API now, so agent-building creators can actually run the comparison this week: GPT-6.1 Terra at $1 per million input tokens against whatever they were paying before — a vendor price cut your audience can verify from their own invoice. And the safety beat is fresh in a way price-cut posts rarely are: per Help Net Security, Astra was designated the first model at this level on September 1, 2026, and Terra now meets the same threshold — which makes the question of what a vendor rating its own model Critical actually obligates it to do answerable from the Preparedness Framework text itself. The system-card addendum ('Addendum to GPT-6 Astra System Card: GPT-6.1 Terra', Publication, October 26, 2026, per the fetched news listing) is the document to open next.

Why it is worth publishing

This card has a quotable number for every creator niche and each one traces to a sentence you can open on screen: the pricing block for AI-business explainers, the 53.6% → 16.7% error-rate drop for AI-quality creators, the Critical designation for the safety beat, and the invoice-entry example for workflow demo channels. The layer discipline is also unusually clean — the official post's numbers and Help Net Security's rendering agree on every figure, so the differentiator is not access but framing: whose measurement, whose price list, whose threshold. The comparisons to Claude Opus 5.5 are OpenAI's claims about a competitor's model, and the card keeps that label on everywhere the number appears.

Evidence basis

Five surface checks this run: the official post opened via the harness reader channel (openai.com bot-gates curl — 403 JS-challenge shell; no Wayback CDX rows), the OpenAI news RSS feed fetched for the pubDate, the openai.com/news listing fetched, Help Net Security opened via raw HTML in full, and the two-channel number cross-check. Calendar (computed once this run from the actual calendar): 2026-10-26 = Monday (both OpenAI posts' dates, RSS pubDates); 2026-10-27 = Tuesday (Help Net Security published_time 2026-10-27T04:33:34+00:00); 2026-09-01 = Tuesday (Astra's Critical designation, per Help Net Security); 2026-10-28 = Wednesday (this card's publish date). Heat: a major vendor pricing-and-capability launch verified against the primary post with an independently opened media cross-check, carrying a fresh safety-governance beat (with the Astra September 1, 2026 precedent, per Help Net Security) — high creator utility across AI-business, AI-quality, and safety-explainer niches.

“OpenAI says its new GPT-6.1 Terra costs half of what the last Terra cost — and OpenAI's own rating puts it at Critical, the higher of its two cyber thresholds.”

Angle

The angle is vendor-claims literacy: a launch post where every number is quotable and every number is the vendor's own — walk the pricing cut, the error-rate drop, and the Critical self-designation as three different KINDS of claim (a list price you can verify from your own invoice, a benchmark only OpenAI ran, a safety rating only OpenAI bestowed).

Format

Carousel

Demo idea

A pricing-and-claims carousel: slide one the official pricing block ($1 / $0.05 / $5 per million tokens, '50% lower' per the post), slide two the error-rate chart (53.6% → 16.7% at low reasoning effort, 4.3% at max) with 'OpenAI reports' on every label, slide three the Critical threshold — Help Net Security's definitional sentence on screen, then the question 'who audits a vendor's own top rating?' with the Astra September 1 precedent as the answer so far.

Platform notes

This card's trust moat is measurement attribution: every benchmark number (15 points, one-third cost, 50.3%, 21.5%, 53.6%, 16.7%, 4.3%, the Opus 5.5 comparisons) is OpenAI-reported — never present one as independently verified, and never round 16.7% into 'one prompt in six' in your own voice (that compression is Help Net Security's rendering; attribute it there or keep the precise figure). The Critical rating is OpenAI's self-designation under its own Preparedness Framework — not an external audit, not a regulator's certification; Help Net Security's definitional sentence ('the higher of two cybersecurity thresholds in the version of the Preparedness Framework dated April 2025') is the citation. The 50%-lower price comparison runs against GPT-5.6 Terra and GPT-6.1 Sol list prices — say the comparison set. The Opus 5.5 comparisons are OpenAI's claims about a competitor's model — Anthropic has not confirmed anything here. The '49.7 percent of tasks unsolved' remainder is Help Net Security's arithmetic, not the vendor's sentence. Availability is per the post's own sentence — Plus, Pro, Enterprise, and Edu in ChatGPT Work and Codex, plus the API as gpt-6.1-terra; no claim beyond those tiers. zh terms stay in English: GPT-6.1 Terra, GPT-5.6 Terra, GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5, Terminal-Bench 4.0, Terminal-Bench-Science 0.1, Agents' Last Exam, Preparedness Framework, Critical, ChatGPT Work, Codex.

Usable claims

OpenAI introduced GPT-6.1 Terra on Monday, October 26, 2026 — per the official post, 'a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing' — with standard API prices the post reports as 50% lower than GPT-5.6 Terra and GPT-6.1 Sol: $1 per million input tokens, $0.05 per million cached input tokens, $5 per million output tokens. Per the same post: on Agents' Last Exam at medium reasoning effort it 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 one-tenth its price per token; at maximum effort the vendor reports it at over three times the token efficiency of Claude Opus 5.5 at one-fourth its price per token — its own arithmetic 'a ~9x lower cost per task' — and over three times faster than both GPT-6 Astra and GPT-6.1 Sol; on an internal benchmark of error-inducing prompts the error rate at low reasoning effort fell from 53.6% (GPT-5.6 Terra) to 16.7%, a drop the post renders as 'nearly 37 percentage points', and at maximum effort it fell 18 points to 4.3%; on Terminal-Bench 4.0 at maximum reasoning effort it solves 50.3% of tasks, 'more than double GPT-5.6 Terra's published score of 21.5%', with the post's own boundary 'For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.'; on Terminal-Bench-Science 0.1 at maximum reasoning effort it surpasses GPT-6.1 Sol at lower cost and comes within five points of Opus 5.5 at less than one-fifth the estimated cost. The vendor reports the model meets the Critical threshold in cybersecurity under its Preparedness Framework, keeps the protections built for GPT-6 Astra, and is more likely than GPT-5.6 Terra to stay within scope and honestly disclose limitations and blockers. Availability per the post: 'Starting today, GPT-6.1 Terra is available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Developers can also access it through the OpenAI API as gpt-6.1-terra.'

Evidence pipeline

From the news

Breakdown

The card carries three kinds of claim, and keeping the kinds apart is the whole skill. Kind one: prices you can verify yourself. The $1 / $0.05 / $5 per-million-token block and the '50% lower' comparison are OpenAI's own list prices against its own prior models — verifiable from your invoice the moment you call the API, which makes them the strongest numbers on the card even though the comparison set is the vendor's own lineup, never a competitor's effective rate. Kind two: benchmarks only OpenAI ran. The Agents' Last Exam gaps (15 points at one-third cost, par with GPT-6 Astra at one-tenth the per-token price), the error-rate drop (53.6% → 16.7% at low reasoning effort, 4.3% at max), Terminal-Bench 4.0's 50.3% against GPT-5.6 Terra's published 21.5%, the Terminal-Bench-Science placement, and every comparison to Claude Opus 5.5 (over three times the token efficiency at one-fourth the price, the ~9x cost-per-task arithmetic, within five points at less than one-fifth the estimated cost) exist only inside OpenAI's own measurement pipeline — Help Net Security's value-add is rendering the same numbers independently, not verifying them, and even its 'one prompt in six' and '49.7 percent of tasks unsolved' are renderings and arithmetic, not new measurements. Kind three: a safety rating only OpenAI bestowed. The Critical rating is a vendor designation under the vendor's own Preparedness Framework; Help Net Security supplies the definitional sentence ('the higher of two cybersecurity thresholds in the version of the Preparedness Framework dated April 2025') and the precedent (Astra, 'its first model at this level', September 1, 2026) — which together let you say on air exactly what the rating is and what it is not: not an audit, not a certification, and in this record a tier with two named members so far — Astra and Terra — whose membership the vendor alone decides. The card's own boundary quotes are load-bearing: 'For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.' is the vendor capping its own benchmark story, and the availability sentence names the exact tiers (Plus, Pro, Enterprise, Edu; ChatGPT Work and Codex; the API as gpt-6.1-terra) — anything beyond those tiers is your invention, not the record's. When a comment quotes the card back at you, the answer is the same in every case: whose sentence says it, and can you open that sentence on screen.

Sources

Risks

On air, keep vendor attribution welded to every number ('OpenAI reports', 'per OpenAI's own post') — never present 50.3%, 16.7%, or the Opus 5.5 comparisons as established facts. Present the Critical rating as what it is: OpenAI's designation of its own model under its own framework — not an external safety certification, and not a claim any regulator or third party has audited. If you compare prices on screen, name the comparison set (GPT-5.6 Terra and GPT-6.1 Sol list prices) and say the numbers are the vendor's. If you mention Claude Opus 5.5 at all, say the comparison is OpenAI's claim about a competitor's model, not Anthropic's.

Demo ideas

A pricing-and-claims carousel: the official pricing block ($1 / $0.05 / $5 per million tokens, '50% lower' per the post), the error-rate chart (53.6% → 16.7% at low reasoning effort, 4.3% at max) with 'OpenAI reports' on every label, then the Critical-threshold slide closing on 'who audits a vendor's own top rating?' with the Astra September 1, 2026 precedent.
An invoice-to-accounting-software workflow demo built on OpenAI's own high-volume example (per Help Net Security): feed a batch of sample invoices, show the entries landing in the accounting sheet, and keep the on-screen disclaimer that the official example is the vendor's own use case — your run is your own measurement.