N° 0The first editionThe monthly · October 2026
A cheaper Haiku; Google’s strongest to defenders first
- 01 · The month in review October: critical PHP flaws need patching; Gemini 4 Argon reached defenders before developers
- 02 · Feature Four models from three makers now list $2 in, $10 out
- 04 · Feature Anthropic's filing shows $974m operating income in June, a $567m loss in September
- 06 · Feature Your AI vendor's incorporation and size decide whether its accounts are public
In this issue
01 to 05 1 / 2 Cover
01 · The month in reviewOctober: critical PHP flaws need patching; Gemini 4 Argon reached defenders before developers
05 · Voice of the weekWho is Sam Altman?Photo: Steve Jennings/Getty Images for TechCrunch / CC BY 2.0, cropped by notis
In this issue
continued, 06 to 11 2of 2
06 · FeatureYour AI vendor's incorporation and size decide whether its accounts are public
07 · FeatureHow to control an AI agent at work, and where each control stops
08 · Explainer · Safety and securityHow prompt injection works and how to limit the damage
09 · Our own recordOctober in our own recordPhoto: Pexels / Pixabay
11 · Back pageMistral releases Large 4 weights, but paid hosting needs a separate licenceLarge 4 weights by month’s end, Mistral said; on 29 October it said it was releasing them
10 · The storiesOctober’s lead stories, week by week
Start with 01October: critical PHP flaws need patching; Gemini 4 Argon reached defenders before developers The month in reviewChecked 5 Nov 20261 of 11
October: critical PHP flaws need patching; Gemini 4 Argon reached defenders before developers
The October review covers the critical PHP patch and the models used to find the flaws, who got new models first, price and agent-control changes, and what compute bills and new rules mean for work.
The one number
69.1%How we countedW3Techs' PHP usage page, opened on 5 November 2026 and dated 4 November 2026, says PHP is used by 69.1% of all the websites whose server-side programming language it knows. Usage statistics and market share of PHP for websites, November 2026, checked 5 Nov 2026.
of sites with a known language use PHP
Illustration made with AI for ai notis
Feature · Models and releasesChecked 4 Nov 20262 of 11
Four models from three makers now list $2 in, $10 out
OpenAI, Anthropic and Google list GPT-6 Sol, GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon at $2 per million input tokens and $10 per million output. Only Google's page dates an end: Argon goes to $4 and $20 on 1 December.
The one number
$2How we countedAnthropic's pricing page lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, opened 4 November 2026. Quote: "Claude Sonnet 5.5 ... $2 / MTok | $10 / MTok" Pricing, checked 4 Nov 2026.
per million tokens · Claude Sonnet 5.5 input, 4 Nov 2026
Illustration made with AI for ai notis
03 · Explained · Agents and apps3 of 11
How prompt caching bills a repeated prompt start
A cached prompt start is billed at a fraction of normal input, but only an identical start counts, and writing to the cache costs more than reading from it.
- A hit needs an identical start: Anthropic says changing tool definitions invalidates the entire cache, and prompts under 512 tokens cannot be cached on Sonnet 5.5 or Opus 5.5.
- Anthropic prices a 5-minute cache write at 1.25 times base input and a read at 0.1 times, or 0.05 times on Opus 5.5 and Sonnet 5.5.
Feature · Money and computeChecked 4 Nov 20264 of 11
Anthropic's filing shows $974m operating income in June, a $567m loss in September
Anthropic's amended S-1, filed on 2 November, adds a September quarter with $16,956 million of revenue and a $567 million operating loss. The filing puts the loss mainly on $5.0 billion of short-term compute and $2.0 billion of non-cash donation-matching charges.
The one number
$974 millionHow we countedThe quarterly table in the S-1 filed on 26 October 2026 states operating income of $974 million for the three months ended 30 June 2026. Anthropic, PBC Form S-1 (filed 26 Oct 2026), checked 4 Nov 2026.
operating income, quarter to 30 Jun 2026
Illustration made with AI for ai notis
05 · Voice of the week5 of 11
Who is Sam Altman?
This week’s Voice of the week profile: the chief executive of OpenAI, who briefed the UN Security Council in September. What he has done and said, what his critics say, and what it means for you.
“It doesn’t matter whether people put the risk of catastrophe at 10%, or 1%, or 12%, or .1%. None of these levels are remotely acceptable.”
“Our vision is simple: we want to create a factory that can produce a gigawatt of new AI infrastructure every week. The execution of this will be extremely difficult; it will take us years to get to this milestone”
Feature · Money and computeChecked 4 Nov 20266 of 11
Your AI vendor's incorporation and size decide whether its accounts are public
We traced where the revenue figures on our list of AI companies come from, then read the filing rules in the UK, Norway and the US, and the terms some vendors' customers sign.
The one number
$113.2 millionHow we countedSynthesia Limited's 2025 group accounts state consolidated revenue of $113.2m for the year ended 31 December 2025. Synthesia Limited group accounts for the year ended 31 December 2025, checked 4 Nov 2026.
Synthesia consolidated revenue, 2025
Illustration made with AI for ai notis
Feature · AI at workChecked 4 Nov 20267 of 11
How to control an AI agent at work, and where each control stops
On 4 November we read what OpenAI, Anthropic, Google, Microsoft and AWS document about an agent's identity, access, spending, stopping and logs, and where each says its own control stops.
The one number
53 centsHow we countedAnthropic's session budgets page gives the example of a session capped at 50 cents that can pause with a list cost of 53 cents, and says this is expected, not a billing error. Session budgets (Managed Agents), checked 4 Nov 2026.
a session capped at 50 cents can reach
Illustration made with AI for ai notis
08 · Explained · Safety and security8 of 11
How prompt injection works and how to limit the damage
OWASP, Anthropic and OpenAI describe prompt injection as untrusted text that a model may follow as an instruction, and the guidance centres on limiting what a fooled model can do.
- OWASP says a prompt injection occurs when prompts alter a model's behaviour in unintended ways, either directly from the user or indirectly through external sources such as websites or files.
- OWASP says it is unclear whether fool-proof prevention exists, and that retrieval augmented generation and fine-tuning "do not fully mitigate" prompt injection vulnerabilities.
Checked · our count
October in our own record
Our count, from our own record, fixed on the day the issue was published.
How we counted
stories published. Stories published in the month, in the daily editions; corrected ones included, withdrawn ones not.
facts checked against their source. Facts an editor checked against their source in the month: confirmed, toned down or left out.
sources opened. Different sources we opened in the month, counted once each however often we opened them.
publishers behind the stories. Different publishers behind the facts in the stories published in the month.
corrections. Corrections made in the month, each shown on its story and on the corrections page.
Stories by section. Stories published in the month, by the section each one sits in.
Hover or focus a number for its own count.
SELECT COUNT(*) FROM notes WHERE form='notice' AND notes.state IN ('published', 'corrected') AND published_at >= :frm AND published_at < :to
stories published
SELECT COUNT(*) FROM claims WHERE from_claim_id IS NULL AND state != 'unchecked' AND checked_at >= :frm AND checked_at < :to
facts checked against their source
2,306 confirmed, 149 toned down, 67 left out
SELECT COUNT(DISTINCT source_id) FROM retrievals WHERE day >= :frm AND day < :to
sources opened
SELECT COUNT(DISTINCT sources.publisher) FROM claims JOIN sources ON sources.id = claims.source_id JOIN notes ON notes.id = claims.note_id WHERE notes.state IN ('published', 'corrected') AND notes.published_at >= :frm AND notes.published_at < :to AND claims.state IN ('agreed', 'narrowed') AND sources.publisher != ''
publishers behind the stories
SELECT COUNT(*) FROM corrections WHERE kind = 'correction' AND at >= :frm AND at < :to
corrections
Go to the back page
Stories by section · 200 in all
Our record13 of 13
Mistral releases Large 4 weights, but paid hosting needs a separate licence
Large 4 weights by month’s end, Mistral said; on 29 October it said it was releasing them
Correction · 3 Oct 2026
OpenAI cuts GPT-6 Sol and Luna API prices against GPT-5.6 promotional rates
Was: GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20. Now: The first key point said GPT-6 Sol was down from $4 and $20. Those were GPT-5.6 Sol's promotional prices, not an earlier GPT-6 Sol price. It now says so.
The monthly · N° 0 · October 2026 · Published 6 Nov 2026. Every fact in it was checked against its source; carried facts say when.