GPT-6.2 Did Not Ship at Cyber at the Frontier: What OpenAI Actually Launched on 12 November

O​penAI's Cyber at the Frontier keynote ran at 10:00 a.m. Pacific on 12 November 2026 and finished without a new model. The model underneath everything announced that morning is G​PT-6 Cyber, which O​penAI had already published on 5 November 2026 at $20 per million input tokens and $100 per million output, behind an access approval most readers will not have; it is priced at exactly twice the rate of GPT-6 Astra, the general-purpose flagship on the same card. G​PT-6.2, the model a large part of the timeline was waiting for, was not mentioned at the keynote, does not appear in O​penAI's model catalogue, its API changelog or its rate card, and has no independent benchmark page. What did launch is two products that sit on top of G​PT-6 Cyber — and unlike a model number, they change something a security team can actually apply for this week.

The two products, and the one model under both

Both announcements are defender tools, both live inside the Daybreak platform, and neither is a model. The distinction matters because the reaction to the keynote — including the post that prompted this piece, which wanted an Apple-style "one more thing" reveal at the end — treated the absence of a model as the story. The products are the more useful half of the morning.

• Daybreak Red Team — early access, by application. Coordinated agents test web applications for exploitable vulnerabilities, validate findings that came from scanners, bug bounties or disclosure reports, and then re-test the patch to see whether the exploit still works. Operators set the scope: which apps and endpoints agents may reach, whether an assessment may use synthetic test data or is limited to read-only checks, how long it runs, and what is excluded. Isolation plus a review layer O​penAI calls Guardian sits under that.

• Daybreak Defense Factory — research preview, installed from the Codex plugin marketplace. Agents take findings from your existing security tooling and from Red Team, trace each one through the code, and produce a patch for an engineer to review and test before merge. It scans a repository once or monitors commits continuously, with a configurable commit-history window and threat model, and keeps a scan history. O​penAI's own developer blog puts a realistic expectation on it: a scan of a large repository "may take several hours."

• The loop between them — Red Team finds and verifies, Defense Factory fixes. That is the actual product thesis, and it is the reason the keynote needed no new weights: the capability being sold is the workflow around a model that shipped a week earlier.

• What is not published — a price. Neither product page carries a rate card, and Red Team's only call to action is a conversation with O​penAI's team about early access. If you need a budget line before you apply, you do not have one yet.

• The commitment behind the programme — O​penAI says it has put $1 billion into subsidised Daybreak access, training and support for utilities, state and local government, community banks, non-profits and open-source projects. That is a vendor statement, not an audited figure, and it is the number to quote if you are arguing for eligibility rather than for tooling.

G​PT-6 Cyber's price card has a step almost nobody quoted

Because the keynote reused a week-old model, the operational news is in the details of that model's rate card rather than in the announcement. G​PT-6 Cyber is billed at $20 input and $100 output per million tokens with cached input at $2 and cache writes at $25 — but only for requests up to 272,000 input tokens. Above that line the same model costs $40 input, $4 cached input and $150 output.

• Price to 272K input — $20.00 per million input, $2.00 cached, $25.00 cache write, $100.00 output.

• Price above 272K input — $40.00 input, $4.00 cached, $150.00 output. The long-context step is roughly double, not a rounding difference.

• Context — 1,050,000 tokens with 128,000 maximum output. The model can hold a repository in one pass; the rate card makes the last three quarters of that window 2x to 1.5x the price of the first quarter.

• Against GPT-6 Astra — $10.00 input and $50.00 output, the same 1,050,000-token window. G​PT-6 Cyber is twice the price at every line, which is the honest way to read a security premium: you are paying for a purpose-trained model plus the vetting around it, not for a bigger one.

• Modalities and limits — text and image in, text out, reasoning set to Highest, speed listed as Fast, no fine-tuning. All of that is O​penAI's own model page; none of it is independently benchmarked, because the cyber tier still has no public evaluation suite that a third party runs.

One practical note for anyone estimating cost from a demo: a full-window pass at 1,050,000 input tokens bills at the above-272K rate for the whole request, so the same document that costs about $21 in the cheap band costs about $42 in the expensive one before a single output token is generated.

The alias in your code is deprecated, and that is the real to-do

The most actionable thing on O​penAI's site this week is not on the keynote page. The Daybreak Red model entry — the alias gpt-daybreak-red-latest that any earlier cyber integration would have been pointed at — resolves to gpt-5.6-cyber and carries a Deprecated badge with the instruction to "use the latest cyber model available to you."

• The alias still works — Live, Chat Completions and Responses endpoints, 400,000-token context, 128,000 max output, a February 2026 knowledge cutoff.

• What it does not do — point at G​PT-6 Cyber. The alias is a fixed pointer to the older model, so a pipeline that has been quietly calling gpt-daybreak-red-latest since the spring is still on G​PT-5.6 Cyber and is missing the 1,050,000-token window the keynote's screenshots assume.

• The migration, if you have the access — move the model id to gpt-6-cyber explicitly rather than trusting an alias, and re-check scope controls and evaluation output, because a 2.6x larger context window changes what an agent will decide to read.

This is the kind of detail a launch-day writeup cannot include, because on 5 November the alias had not yet been stamped. It is the strongest argument for reading the docs a week late.

Why "G​PT-6.2" keeps appearing in searches for a model that does not exist

On 11 November the Codex product lead posted that Thursday was "feeling codexy" and might be "a good time for a banked reset" — a remark about subscription quota. It was quote-posted within four minutes as a launch notice for G​PT-6.2, and by the following evening the keynote's silence was being read as a delay of something that had never been announced. The same expectation shows up in the post this article started from, which wanted one more thing at the end of the keynote.

The catalogue does not support it. O​penAI's model index and price list carry five G​PT-6-family identifiers — GPT-6 Astra, GPT-6.1 Sol, G​PT-6.1 Terra, GPT-6 Luna and G​PT-6 Cyber — and not one of them contains "6.2". The API changelog's most recent model entries are G​PT-6 Cyber on 5 November and the two GPT-Image-2.6 models on 4 November. Artificial Analysis, which indexes new models within days, has no page for any 6.2 variant. A reset is a coupon on a ChatGPT or Codex plan; it is not a model, and the two have been conflated for two days running.

There is a version of this where the disappointment is about Apple rather than O​penAI. The post's first clause reaches for an Apple-style surprise, and Apple's own 13 November news was hardware — the fourth-generation Apple TV 4K at $299 with an A19 chip, and the HomeView hub at the same price, both arriving in Australia and New Zealand first, with Siri AI rolling back to older Apple TV 4K boxes in tvOS 27.2.1. Apple's actual model family, AFM 3, dates from September 2026. Nothing in either company's week was a frontier model launch, which is why the week reads as flat even though two shipping products landed in it.

What a team can do with this on 13 November

Three things are available now and one is not. Red Team takes applications; Defense Factory ships as a research preview you install from the Codex plugin marketplace, which makes it the cheaper way to find out whether the agent loop is worth pursuing before you ask anyone for Red Team access; and the older cyber model remains callable through the deprecated alias for organisations already approved. G​PT-6.2 is not on any of those lists.

If the reason you wanted a new model was general capability rather than cyber work, the models that shipped earlier in this generation are callable today through one OrcaRouter key, at provider list price with nothing added on top — GPT-6 Astra at $10.00 / $50.00, GPT-6.1 Sol at $2.00 / $10.00, GPT-6 Luna at $0.10 / $0.50. Because the platform passes the provider's list price straight through, an O​penAI price change is live on our side the same day it lands rather than at the next contract renewal, and automatic failover means one provider's capacity problem does not take the route down with it. G​PT-6 Cyber is not in that catalogue: it is gated behind Daybreak Red approval at O​penAI, so no router can hand it to you, and any page claiming otherwise is describing something else.

For teams whose work is genuinely offensive-security research, the honest position is that access is the constraint this month, not model quality. The application is the bottleneck, the products on top are previews without published prices, and the only number you can plan against is the rate card — $20 / $100 below 272K input tokens, $40 / $150 above it, with a deprecated alias in the path until you change it.