The record for this issue
Held at the gate.
114 claims in the draft, each tested by a second agent against the page it cites. 114 held. Nothing had to be cut.
-
01
titleconfirmed
Held at the gate. (models held back before release)
The page says “OpenAI scrapped the release of GPT-6.1 Astra ... planned for an October debut”
-
02
dekconfirmed
OpenAI pulled a new model a day before DevDay
The page says “announcement came a day before ... conference on Tuesday in San Francisco (Express & Star, 29 Sep 2026); Fast Company: DevDay Tuesday, Sept. 29”
-
03
dekconfirmed
...because it didn't stay within its permissions
The page says “"didn't quite meet the bar in terms of staying within scope and authorisation"”
-
04
dekconfirmed
then launched agents that run around the clock
The page says “"always-on agents" ... "can work towards your goals 24/7"”
-
05
dekconfirmed
Google held its own new model back too
The page says “rolling out to a set of trusted cyber defenders through our Fairwind Program ... before wider release”
-
06
dekconfirmed
Who gets a model, and when, became a safety control (highlight)
The page says “Google describes this as a phased approach to releasing frontier capabilities”
-
07
summaryconfirmed
OpenAI shelved GPT-6.1 Astra
The page says “OpenAI scraps new AI model release over safety concerns”
-
08
summaryconfirmed
and launched dots
The page says “Introducing dots, published September 29, 2026”
-
09
summaryconfirmed
Google limited Gemini 4 Argon to cyber defenders
The page says “rolling out to a set of trusted cyber defenders through our Fairwind Program”
-
10
summaryconfirmed
six AI companies signed a voluntary White House accord
The page says “signed at a White House meeting on Tuesday, September 29 ... Amodei, Pichai, Zuckerberg, Brockman, Huang, Musk ... 'voluntary accord'”
-
11
art alt / art svgconfirmed
Yellow ball stopped at a barrier labelled HELD; SVG text 'HELD'
The page says “Illustrative; word HELD matches the Astra hold”
-
12
The week in four, item 1confirmed
A model got pulled.
The page says “OpenAI scrapped the release of GPT-6.1 Astra”
-
13
The week in four, item 1confirmed
GPT-6.1 Astra's October release
The page says “"a next-generation AI model planned for an October debut"”
-
14
The week in four, item 1confirmed
called off on Sept. 28
The page says “Express & Star (09:03 BST 29 Sep) says the announcement came a day before executives met Trump (Tue 29 Sep); Fast Company: 'on Monday said it was holding off'”
-
15
The week in four, item 1confirmed
its safety head said it fell short on staying within scope and authorization
The page says “Saachi Jain, head of safety systems ... "didn't quite meet the bar in terms of staying within scope and authorisation"”
-
16
The week in four, item 2confirmed
At DevDay, OpenAI launched dots
The page says “OpenAI DevDay 2026 conference, Tuesday, Sept. 29 ... Dots are OpenAI's new agents”
-
17
The week in four, item 2confirmed
ChatGPT agents with their own cloud computer
The page says “"Dots can do nearly anything using their own cloud computer, their own browser, and the apps you've connected"”
-
18
The week in four, item 2confirmed
keep working toward your goals when you're away
The page says “"can work towards your goals 24/7"; when you aren't actively working with yours, it "looks for ways to help in the background"”
-
19
The week in four, item 3confirmed
Google released Gemini 4 Argon to vetted cyber defenders only
The page says “"rolling out to a set of trusted cyber defenders through our Fairwind Program"”
-
20
The week in four, item 3confirmed
while it finishes testing the guardrails for everyone else
The page says “gather feedback from early testers and iterate on guardrails before wider release”
-
21
The week in four, item 4confirmed
Six AI companies signed a voluntary White House accord
The page says “'voluntary accord' signed at a White House meeting by leaders of Anthropic, Google, Meta, OpenAI, Nvidia, xAI”
-
22
The week in four, item 4confirmed
the FTC confirmed an investigation
The page says “An FTC spokesperson confirmed the investigation but declined further comment”
-
23
The week in four, item 4confirmed
California barred firings decided by AI alone
The page says “SB 947 ... bars employers from relying only on AI when making a disciplinary action or termination decision”
-
24
The storyconfirmed
On Monday, Sept. 28, a day before its developer conference
The page says “on Monday said it was holding off on releasing it ... DevDay 2026 conference, Tuesday, Sept. 29, 2026, in San Francisco”
-
25
The storyconfirmed
called off the planned October release of GPT-6.1 Astra
The page says “"planned for an October debut"”
-
26
The storyconfirmed
Saachi Jain, its head of safety systems
The page says “"head of safety systems at OpenAI"”
-
27
The storyconfirmed
Quote: "didn't quite meet the bar in terms of staying within scope and authorization." (13 words, attributed)
The page says “"didn't quite meet the bar in terms of staying within scope and authorisation"”
-
28
The storyconfirmed
AAP reported internal testing found it did not meet safety and alignment standards
The page says “"after internal testing found the system did not meet the company's safety and alignment standards"”
-
29
The storyconfirmed
Training of OpenAI's most advanced models was already paused
The page says “OpenAI "paused training of its most advanced models last week"”
-
30
The storyconfirmed
The next day, at DevDay in San Francisco
The page says “OpenAI DevDay 2026 conference, Tuesday, Sept. 29, 2026, in San Francisco”
-
31
The storyconfirmed
dots: always-on agents in ChatGPT
The page says “"remarkably capable, always-on agents" ... reach yours in ChatGPT, Slack, or Teams”
-
32
The storyconfirmed
each with its own cloud computer and browser
The page says “"their own cloud computer, their own browser"”
-
33
The storyconfirmed
running on the current GPT-6 Astra
The page says “"Powered by GPT-6 Astra"”
-
34
The storyconfirmed
able to connect to more than 4,000 apps
The page says “"can readily connect to over 4,000 apps"”
-
35
The storyconfirmed
when you aren't working with one, a dot does proactive research through read-only tools
The page says “Proactive research uses tools "restricted to be read-only"”
-
36
The storyconfirmed
so it can't send messages, change what's in your apps or control your computer
The page says “"they can't send messages, change app content, or control your browser or computer"”
-
37
The storyconfirmed
Custom Rules let you allow an action, require approval for it or block it
The page says “"Custom Rules let you allow specific actions, require approval, or block them"”
-
38
The storyconfirmed
Google drew its line on Sept. 30 (Argon date)
The page says “Publication date: Sep 30, 2026”
-
39
The storyconfirmed
Argon can find, validate and patch critical software vulnerabilities on its own
The page says “"can autonomously find, validate, and patch critical software vulnerabilities"”
-
40
The storyconfirmed
went first to "trusted cyber defenders" in its Fairwind Program
The page says “"rolling out to a set of trusted cyber defenders through our Fairwind Program"”
-
41
The storyconfirmed
monitors watch the model's reasoning and actions and can stop it
The page says “"monitor Argon's chain-of-thought and actions and stop execution when necessary"”
-
42
The storyconfirmed
wider access, starting with paid API customers and Google AI Ultra subscribers, waits until guardrail testing is done
The page says “iterate on guardrails before wider release ... "starting with paid API customers and Google AI Ultra subscribers"”
-
43
The storyconfirmed
the test Astra failed is staying inside permissions; dots answer with read-only defaults and rules; Google by choosing who goes first
The page says “read-only proactive research + Custom Rules; Argon to trusted defenders first (Google); Astra short on 'scope and authorisation' (AAP)”
-
44
The storyconfirmed
Three days of holding back and shipping; Dates as reported, spacing not to scale
The page says “Events span Sept. 28-30”
-
45
The story, step 1confirmed
Sept. 28: OpenAI pulls GPT-6.1 Astra's October release
The page says “"on Monday said it was holding off" (Monday = Sept. 28); AAP: 'planned for an October debut'”
-
46
The story, step 2confirmed
Sept. 29: DevDay: dots launch
The page says “Publication date: September 29, 2026”
-
47
The story, step 2confirmed
Sept. 29: GPT-6.1 Sol launches at DevDay
The page says “conference Tuesday, Sept. 29 ... OpenAI announced GPT-6.1 Sol”
-
48
The story, step 3confirmed
Sept. 29: Six AI companies sign a voluntary White House accord
The page says “signed at a White House meeting on Tuesday, September 29, 2026”
-
49
The story, step 4confirmed
Sept. 30: Gemini 4 Argon goes to cyber defenders first
The page says “Sep 30, 2026”
-
50
The story, step 5confirmed
Sept. 30: FTC confirms its AI investigation
The page says “The agency announced it on Wednesday [Sept. 30]; an FTC spokesperson confirmed the investigation”
-
51
The story, step 5confirmed
Sept. 30: Newsom signs 13 AI bills
The page says “Sep 30, 2026; lists 13 bills signed (TNW: 'Newsom signed 13 AI bills')”
-
52
The storyconfirmed
What a dot may do on its own; OpenAI's controls
The page says “Custom Rules / built-in safety requirements”
-
53
The story, item 1confirmed
Built-in safety requirements, always applied
The page says “"Built-in safety requirements always apply"”
-
54
The story, item 2confirmed
Read-only tools while you're away
The page says “proactive research uses tools "restricted to be read-only" when you aren't actively working with it”
- 55 confirmed
-
56
The story, item 4confirmed
Custom Rules: block
The page says “"or block them"”
-
57
The storyconfirmed
When you aren't working with a dot it uses read-only tools: can't send messages, change app content or control your computer; Custom Rules decide allow/approve/block
The page says “"they can't send messages, change app content, or control your browser or computer"; "Custom Rules let you allow specific actions, require approval, or block them"”
-
58
The storyconfirmed
The model that overstepped stayed in the lab. The agents that might shipped with a rulebook.
The page says “OpenAI had to weigh [persistence] against unauthorised behaviour; Astra not released”
-
59
The storyconfirmed
One model held, one agent shipped.
The page says “Astra held (AAP); dots launched Sept. 29”
-
60
shipped, item 1confirmed
OpenAI · GPT-6.1 Sol · Sept. 29
The page says “OpenAI DevDay 2026, Tuesday, Sept. 29 ... OpenAI announced GPT-6.1 Sol”
-
61
shipped, item 1confirmed
A cheaper model for agents and code
The page says “"Near-Astra intelligence for a fifth of the price"; cost-effective for agentic coding, computer use”
-
62
shipped, item 1confirmed
In ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users
The page says “"available starting today to all Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex"”
-
63
shipped, item 1confirmed
$2 per million input tokens and $10 per million output
The page says “$2 per million input tokens and $10 per million output tokens”
-
64
shipped, item 1confirmed
one-fifth of Astra's standard prices
The page says “Astra standard $10 input / $50 output; Sol's are one-fifth”
-
65
shipped, item 1confirmed
A faster Ultrafast version is due "in the coming days."
The page says “GPT-6.1 Sol Ultrafast will launch "in the coming days"”
-
66
shipped, item 2confirmed
Anthropic · Claude Sonnet 5.5 · Sept. 28
The page says “Release date: September 28, 2026”
-
67
shipped, item 2confirmed
much better at agent work
The page says “Terminal-Bench 4.0 70.6% vs 10.3%; OSWorld 2.1 80.1% vs 57.0%”
-
68
shipped, item 2confirmed
A faster, cheaper companion to Opus 5.5
The page says “"a faster, cheaper companion to Opus 5.5"”
-
69
shipped, item 2confirmed
for scoped tasks, bug fixes and documents
The page says “"well-defined everyday work, such as fixing bugs and producing polished documents, slides, and spreadsheets"”
-
70
shipped, item 2confirmed
Same $2/$10 per million tokens as Sonnet 5
The page says “Input: $2, the same as Sonnet 5. Output: $10, the same as Sonnet 5”
-
71
shipped, item 2confirmed
Anthropic reports large gains on agentic coding and computer use
The page says “Terminal-Bench 4.0 far above Sonnet 5; OSWorld 2.1 large gain over Sonnet 5”
-
72
shipped, item 3confirmed
Google · Skills in Gemini · Sept. 30
The page says “Publication date: September 30, 2026”
-
73
shipped, item 3confirmed
Skills are saved instructions that work inline in a chat
The page says “Reusable custom instructions ... Skills work inline in your chat threads”
-
74
shipped, item 3confirmed
Reusable instructions you can stack
The page says “you can stack several in one prompt”
-
75
shipped, item 3confirmed
they use the open SKILL.md format, so skills made elsewhere carry over
The page says “open, Markdown-based SKILL.md standard, so you can copy skills created on other platforms”
-
76
shipped, item 3confirmed
Google said Workspace rollout starts Oct. 5
The page says “Workspace, Rapid Release domains: rollout ... October 5, 2026”
-
77
shipped, item 3confirmed
Gems keep working for now
The page says “Users can still create and use Gems in the Gemini app”
-
78
shippedconfirmed
Terminal-Bench 4.0 (agentic coding): Sonnet 5.5 70.6%, Sonnet 5 10.3%
The page says “Terminal-Bench 4.0 | 70.6% | 10.3%”
-
79
shippedconfirmed
OSWorld 2.1 (computer use): Sonnet 5.5 80.1%, Sonnet 5 57.0%
The page says “OSWorld 2.1 (partial) | 80.1% | 57.0%”
-
80
shippedconfirmed
Sonnet 5.5 against Sonnet 5; Anthropic's own benchmark results; vendor-reported, not independently verified
The page says “These figures are Anthropic's own reported results”
-
81
shippedconfirmed
Sol, Sonnet 5.5 and Gemini 4 Argon's introductory rate all list at $2 in and $10 out per million tokens
The page says “Argon introductory rate $2/$10 (later $4/$20); Sol $2/$10 (OpenAI); Sonnet 5.5 $2/$10 (Anthropic)”
-
82
shippedconfirmed
Faster, cheaper, reusable.
The page says “Ultrafast / one-fifth price / reusable skills”
-
83
The storyconfirmed
On Sept. 29 leaders of Anthropic, Google, Meta, Nvidia, OpenAI and xAI signed a voluntary accord with President Trump
The page says “signed by Trump, Amodei, Pichai, Zuckerberg, Brockman, Huang, Musk ... White House meeting on Tuesday, September 29”
-
84
The storyconfirmed
pledged "robust internal controls"
The page says “companies would implement "robust internal controls"”
-
85
The storyconfirmed
an "independent external auditor" to check them
The page says “partner with an "independent external auditor" to assess whether those controls work”
-
86
The storyconfirmed
a board committee to review the audits
The page says “calls for a board-level committee at each company to review internal and external audit reports”
-
87
The storyconfirmed
accord says it "may make sense" to write the steps into law over time
The page says “"Over time, it may make sense to codify these steps into laws and regulations."”
-
88
The storyconfirmed
The President said he saw "tremendous self-policing"
The page says “"I think I'm seeing tremendous self-policing."”
-
89
The storyconfirmed
A day later, an FTC spokesperson confirmed an investigation
The page says “Wednesday 30 September 2026 ... An FTC spokesperson confirmed the investigation”
-
90
The storyconfirmed
into OpenAI, Anthropic and other AI companies over risks to consumers
The page says “OpenAI and Anthropic, plus "other artificial intelligence companies" ... "dangers their technology may pose to consumers"”
-
91
The storyconfirmed
California's governor signed 13 AI bills (Sept. 30)
The page says “Newsom signed 13 AI bills ... Wednesday, September 30”
-
92
The storyconfirmed
SB 947 bars employers from relying on an automated system alone to discipline or fire workers
The page says “Employers cannot rely only on an automated decision system to discipline or fire workers”
-
93
The storyconfirmed
takes effect July 1, 2027
The page says “Effective date: July 1, 2027”
-
94
The storyconfirmed
Governor: AI should expand opportunity, "not come at the expense of workers and families."
The page says “"AI should expand opportunity – not come at the expense of workers and families."”
-
95
The storyconfirmed
A pledge, a probe and a person in the loop.
The page says “SB 947 requires human review”
-
96
Companies now stage who gets a modelconfirmed
OpenAI held a model, shipped agents with read-only defaults, Google picked who goes first
The page says “read-only proactive research; Argon to trusted defenders first”
-
97
What to watchconfirmed
Coming up · Oct. 5 – Oct. 14
The page says “Items run Oct. 5 to Oct. 14”
-
98
What to watch, item 1confirmed
Mon 10/5: Rapid Release domains get skills from today, finishing by Oct. 12
The page says “Rapid Release: rollout October 5, 2026, completion expected by October 12, 2026 (announced Sept. 30)”
-
99
What to watch, item 2confirmed
Sol Ultrafast arrives in the coming days, up to 8x faster output in Codex
The page says “"in the coming days," with "up to 8x faster token generation" than standard speed in Codex”
-
100
What to watch, item 3confirmed
Training pause: no restart date; safeguards it said must come first; no new Astra date
The page says “training would resume "only when we are confident that we have additional safeguards"; Express & Star gives no new release date”
-
101
What to watch, item 4confirmed
Formal civil investigative demands expected, per TNW
The page says “"the agency is drafting civil investigative demands" ... "the FTC expects to send the demands in the coming weeks"”
-
102
What to watch, item 5confirmed
Wed 10/14: GPT-5.5 leaves Codex for users signed in with ChatGPT, at noon Pacific
The page says “"Effective October 14, 2026, at noon PDT (19:00 UTC), GPT-5.5 retires from Codex for users signed in with ChatGPT."”
-
103
The checklist, item 3confirmed
New defaults and retirements arrive weekly
The page says “This week: Sol launch (OpenAI Sol) and GPT-5.5 retirement notice”
-
104
The checklist, item 4confirmed
In California, employers won't be able to discipline or fire on an automated system's output alone
The page says “Employers cannot rely only on an automated decision system to discipline or fire workers ... effective July 1, 2027”
-
105
quoteconfirmed
"We have an extremely high bar in terms of safety and alignment." Saachi Jain, via AAP (11 words)
The page says “"But when we ship it to users, we have an extremely high bar in terms of safety and alignment."”
-
106
Before you go, item 1confirmed
OpenAI's newest model was held back for not staying in scope
The page says “"didn't quite meet the bar in terms of staying within scope and authorisation"”
-
107
Before you go, item 2confirmed
Dots work 24/7 and can't send messages while you're away
The page says “"can work towards your goals 24/7"; read-only tools "can't send messages"”
-
108
Before you go, item 3confirmed
Three new models, three companies, one price: $2 in, $10 out
The page says “Argon intro $2/$10; Sol $2/$10; Sonnet 5.5 $2/$10”
-
109
linkedin postconfirmed
OpenAI pulled GPT-6.1 Astra a day before its developer conference
The page says “on Monday said it was holding off; DevDay Tuesday, Sept. 29”
-
110
linkedin postconfirmed
Its safety head said it fell short on staying within scope and authorization
The page says “"didn't quite meet the bar in terms of staying within scope and authorisation"”
-
111
linkedin postconfirmed
The next day the same company launched dots, agents that keep working on your goals while you're away
The page says “"can work towards your goals 24/7"; works in the background”
-
112
linkedin postconfirmed
They're read-only in the background, and you write the rules
The page says “proactive research tools "restricted to be read-only"; Custom Rules”
-
113
linkedin postconfirmed
Google ... Gemini 4 Argon, giving it to vetted cyber defenders first
The page says “trusted cyber defenders through our Fairwind Program”
-
114
linkedin postconfirmed
Issue two of Predictive Text; the new $2/$10 model price point
The page says “number: 2 in draft; $2/$10 for Sol, Sonnet 5.5, Argon intro”
The checker's notes
- Back issue dated 2026-10-05. All cited facts were published Sept. 28 - Oct. 2, 2026, inside the 8-day window and before D; nothing requires knowledge after D.
- The training pause (week before Sept. 29) predates the window; the text frames it as 'already paused', which reads as background.
- Non-blocking src fixes: add Fast Company to the timeline src and the Sol shipped item src (OpenAI Sol page shows no date; AAP gives no date for Sept. 28); add src fields to the ahead items for TNW (FTC) and OpenAI models.
- Post-D development (not a window issue): Google's Oct. 13 update says the Workspace skills launch was delayed.
- No securities/markets content; no first-person statements; quotes are 13 and 11 words, attributed.