← Back to articlesAI news
Alexandre8 min read
AI News November 6, 2026: Anthropic at $17 Billion a Quarter, OpenAI Cuts Its Raise and ChatGPT at 20% of Google
Anthropic reports $17 billion in Q3 revenue, OpenAI cuts its raise to $20 billion, Meta borrows at higher yields and ChatGPT claims 20% of Google's query volume.
Alexandre8 min read
Credit: CNBC / AFP
Anthropic reported $17.0 billion in revenue for the third quarter of 2026, up from $1.4 billion a year earlier, in the amended version of its S-1 filing. The same document puts its purchase and lease commitments at $895.4 billion, almost entirely devoted to compute and infrastructure.
Meanwhile, OpenAI cuts its fundraising target from $30 billion to $20 billion, Meta borrows on the bond market at higher yields than in April, and OpenAI says search queries going through ChatGPT reach about 20% of Google's volume. Google, for its part, cuts off Search for schools that refuse its AI tools.
Breakdown.
ChatGPT claims 20% of Google's search volume
OpenAI says queries that trigger a web search in ChatGPT amount to about 20% of Google's volume, with 1.2 billion weekly users, up from 1 billion sixty days earlier. The figure is an OpenAI estimate announced on November 2, with no independent source for Google's volume.
Click to enlarge
Credit: The Verge
Adam Fry, head of consumer product for ChatGPT, presented the numbers at a press event, The Verge reports. The metric counts any query that triggers a web search, not only queries comparable to a classic search. Fry gave more detail to Big Technology, without revealing where the Google reference volume comes from.
Query volume does not convert mechanically into revenue: Google still earns most of its advertising income from Search. The shift toward interfaces where you ask a question instead of browsing a list of links was already described in the October 24 AI news.
Google has just responded on a field where it holds a distribution advantage: schools.
Google turns off Search for schools that refuse its AI tools
Google will turn off Gemini, Gemini Notebook, Gemini in Classroom and Search on K-12 student accounts in the United States unless administrators explicitly opt in. A school that refuses AI tools therefore also loses Search, according to Google statements to the New York Times.
Click to enlarge
Credit: Ars Technica
Chris Phillips, vice president of education at Google, justifies the choice by the deep integration of search and AI. Schools will be able to choose standard Gemini or a "Guided Learning" mode, designed to guide students without handing them the answer. Google starts testing this mode in a selection of US schools this month (Ars Technica, The Verge).
Search acts as leverage here: a feature almost no school can remove is tied to another one that was optional until now. The bundling recalls antitrust debates on tied sales, though no proceeding has been mentioned at this stage.
While platforms tie their products together, the web's infrastructure is trying to settle another question: who has the right to train on published content.
Google, Microsoft, OpenAI and Anthropic back a single robots.txt signal against training
On October 30 and 31, Google, Microsoft, OpenAI and Anthropic announced their intention to honor a single signal in the robots.txt file to opt out of model training. A draft has been filed with the IETF, the body that maintains the robots exclusion protocol, with Microsoft and OpenAI in the lead.
Click to enlarge
Credit: Coywolf
Until now, a publisher had to write one rule per crawler: GPTBot, ClaudeBot, Google-Extended, and so on. The draft, titled "Machine-Readable Training Preference Signal" and written by Kevin Kelley, technical standards lead at Anthropic within the IETF's AI Preferences working group, defines a single token that participating providers would read as the site's preference (Search Engine Journal). Microsoft and OpenAI plan to implement it in the coming months, Google next year.
The limits are documented: it is an individual draft, not an approved standard, and honoring a robots.txt preference remains voluntary (Coywolf). Legal avenues (licenses, copyright) do not change.
On the numbers side, the question of content value weighs all the more as the amounts committed to training reach unprecedented levels, as the financial documents published this week show.
AI financing: Anthropic's S-1, OpenAI's reduced raise and Meta's borrowing
Anthropic reports $17.0 billion in revenue for the third quarter of 2026 and $895.4 billion in compute commitments. OpenAI cuts its raise to $20 billion at a $1.25 trillion valuation. Meta borrows at least $15 billion in bonds, at yields above those of April.
Click to enlarge
Credit: Tech Funding News
Anthropic: $17 billion in a quarter, $895 billion in commitments
On November 2, Anthropic filed an amended version of its S-1 (the registration document a company submits to the SEC before an IPO). The initial version covered only the first half. According to CNBC, the third quarter shows about $17 billion in revenue and a net loss of $23.29 billion. Over nine months, revenue reaches $33.24 billion, about 1,200% more than a year earlier, for a net loss of $104 billion, largely tied to fair-value changes on convertible notes and derivatives.
The S-1/A filed with the SEC details the mechanics: an operating margin of minus 3% (operating loss of $567 million) versus minus 224% a year earlier, an adjusted operating margin of 15% ($2.5 billion) and more than 11,900 customers above $100,000 in trailing twelve-month revenue, versus about 400 in March 2025. Total purchase and lease commitments reach about $895.4 billion, "substantially all" of it for compute and infrastructure.
The first filing, on October 26, was covered in the November 1 AI news. Price, share count and valuation remain to be set.
OpenAI: $20 billion targeted instead of $30 billion
On October 30, the Financial Times reported that OpenAI was cutting its raise to about $20 billion, from a $30 billion target, at a $1.25 trillion valuation, below the $1.4 trillion floated at the end of September (Tech Funding News). The Wall Street Journal says commitments gathered are around $11 billion, less than half the initial goal (WSJ).
A United Arab Emirates consortium, including the MGX fund, is expected to be the first participant, with an option to rent compute capacity in the region (Forbes). According to Bloomberg, OpenAI's annualized revenue (the run-rate, meaning the last month's revenue multiplied by twelve) was approaching $50 billion at the end of September. At $1.25 trillion, the valuation is about 25 times that figure. The raise comes ahead of Anthropic's IPO, expected in mid-November according to several outlets.
Meta: a bond sale at higher yields
Click to enlarge
Credit: The Next Web
Meta is looking to raise at least $20 billion in bonds, including at least $15 billion in dollars and a first euro issue, according to Bloomberg relayed by The Next Web. The 30-year bonds are being discussed at around 2.5 points above US Treasuries, nearly 0.9 points more than in April's issue.
Markets are reacting: Big Tech bonds trade at the widest gap with the rest of investment-grade credit since 2003, after nearly $430 billion issued since last year. S&P revised Meta's outlook to negative, with the AA- rating maintained. Meta plans $135 to $145 billion in capital expenditure in 2026 and reports $825 billion in lease and contract commitments, according to Quartz.
Part of these commitments now shows up in user billing: OpenAI announced an overhaul of its usage limits for December.
ChatGPT and Codex will share a single token quota from December
Starting in December, OpenAI's Plus, Pro and Business plans will draw from a single token quota covering ChatGPT and Codex. Instant chat stays outside the quota. The change, published in OpenAI's help center on October 31, makes agent usage the unit of measure for consumer plans.
Click to enlarge
Credit: The Decoder
A token is the fragment of text (about three quarters of a word) that models read and produce, and the unit APIs bill on. According to the OpenAI help center article, consumption depends on the model chosen and the task requested. Until now, conversation and code agents had separate limits.
For developers, the practical consequence is to measure consumption before December. The blog has documented this measurement work on the Claude Code side (how I cut my Claude Code token usage in half), and the agent architecture explains why a refactoring session can weigh more than a week of conversation (architecture of AI code agents).
Other news in brief
CrowdStrike acquires Fiddler AI and Dreadnode: two startups specialized in securing AI agents, for about $495 million according to amounts reported by the specialized press.
Fal doubles its valuation to $16 billion: the inference platform raised $1.3 billion, led by Bessemer, with Nvidia in the round. It reports $800 million in annualized revenue and had to limit customer access this summer for lack of capacity (TechCrunch).
GPT-6 Astra designed a hardware device in under a week: OpenAI says the model produced the enclosure, circuit board, firmware and interface, with no direct human modification. Several revision cycles were needed because manufacturing tolerances were not accounted for (The Decoder).
Stanford: 44% of cited claims about LLMs are outdated: retesting 50 claims on 20 models released between July 2024 and September 2026, Owen Queen and James Zou find an average shelf life of 15.7 months (Stanford).
Brookfield and Drax target a data center of more than 1 gigawatt: investment estimated at more than 20 billion pounds ($27 billion), with the first 100 megawatts targeted for early 2028 (Drax and Brookfield press release).
OpenAI hires former Army Secretary Dan Driscoll: as senior adviser for national security modernization (CNBC).
OpenAI prepares for a possible IPO in 2027: president Greg Brockman says the company is laying the regulatory and accounting groundwork to be ready by that horizon (TechCrunch).
Anthropic launches Personal Finance in Claude: the tool analyzes bank statements and categorizes spending, relying on MCP and tool calling (TechCrunch).
Conclusion: AI is now financed in public
The figures published this week come from regulatory filings and bond markets more than from product demos. An S-1 forces a company to put numbers on its commitments. A bond sale reveals the price investors demand. A reduced raise shows where the appetite sits.
At the same time, usage keeps growing: 1.2 billion weekly users for ChatGPT, more than 11,900 customers above $100,000 in revenue for Anthropic. The next public milestone is Anthropic's IPO, expected in mid-November according to several outlets. OpenAI, for its part, talks about preparing for 2027.
Alex
- Anthropic reports $17.0 billion in Q3 2026 revenue and $895.4 billion in compute commitments
- OpenAI targets $20 billion instead of $30 billion, at a $1.25 trillion valuation
- Meta borrows at least $15 billion in bonds, with wider spreads than in April
- Google, Microsoft, OpenAI and Anthropic back a single robots.txt signal against training
- ChatGPT and Codex will share a single token quota from December
Related articles
AI News November 1, 2026: Anthropic Files Its S-1, Gemini 4 Carbon and Mistral Large 4
Anthropic reveals $16.3 billion in revenue in its IPO S-1, Google launches Gemini 4 Carbon, Mistral opens Large 4 and the Nasdaq loses $2 trillion.
AI News October 24, 2026: Anthropic's Claude Duet, Google's Agent Compute Engine, OpenAI on Google Cloud
Anthropic launches Claude Duet, Google unveils Agent Compute Engine, OpenAI lands on Google Cloud, and Muse passes 10 million users.
AI News October 16, 2026: Anthropic books $16.8B in a quarter, Gemini 4 Argon at $2, Character.ai goes bankrupt
Anthropic reports $16.8B in Q3 ahead of its IPO, Gemini 4 Argon ships at $2, Meta teams up with Apple, Microsoft takes on OpenAI and Character.ai files Chapter 11.Comments
Comments
Got a take on this article?
Create a free account in 10 seconds to comment, like, and get the next articles straight to your inbox.
Don't have an account yet?