← Back to articlesAI news
Alexandre11 min read
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.
Alexandre11 min read
Credit: Reuters, via CNBC
Anthropic brought in $16.8 billion in revenue between July and September. The figure was presented on Wednesday, October 14, in San Francisco, at an investor day held one month before an initial public offering planned for November. It represents a 45% increase in a single quarter, and yet some investors are revising the valuation downward.
Google makes Gemini 4 Argon public at $2 per million tokens. Meta ships its first frontier model, then runs an 816-billion-parameter open-weight model on Mac Studios, with Apple's blessing. Microsoft claims it beat OpenAI with a quarter of the compute. And Character.ai files for bankruptcy while eight US states take OpenAI to court.
Breakdown.
Click to enlarge
Gemini 4 Argon lands at Sonnet 5.5 pricing
Google opened Gemini 4 Argon to the public on Tuesday, October 13, for Pro and Ultra subscribers and through the API. The launch price, $2 per million input tokens and $10 per million output tokens until November 30, puts the model level with Sonnet 5.5 while it matches GPT-6 Astra on the Artificial Analysis leaderboard.
The price has an expiration date. Starting December 1, Argon moves to $4 for input and $20 for output (VentureBeat). Google adds a 95% discount on cached tokens, or $0.10 per million. Prompt caching covers tokens already sent (a long context, a codebase, documentation) that the model reuses from one request to the next without reprocessing them. For a coding agent that rereads the same repository fifty times a day, that is the line that weighs most on the bill.
Another concrete change: the output limit goes from 64,000 to 1 million tokens. A single call can now produce the equivalent of several books.
Here is where Argon sits in the current price grid, in dollars per million tokens:
| Model | Input | Output |
|---|---|---|
| Gemini 4 Argon (until November 30) | $2 | $10 |
| Gemini 4 Argon (from December 1) | $4 | $20 |
| Claude Sonnet 5.5 | $2 | $10 |
| GPT-6.1 Sol | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
| Muse Wonder | $4 | $20 |
| GPT-6 Astra | $10 | $50 |
Benchmarks aimed at agent work
On DeepSWE v1.1, a software engineering task benchmark, Argon reaches 77.9%, ahead of Opus 5.5 (74.2%) and GPT-6 Astra (74.1%). On AutomationBench, the gap widens: 51.3% versus 42.5% for Opus. The clearest score concerns Harvey Legal, a legal work test: 19.6% for Argon, 5.4% for Astra, 3.8% for Opus.
Opus 5.5 keeps the edge on Terminal-Bench 4.0, with 66.4% versus 57.4%. And on the Artificial Analysis Intelligence Index, Argon scores 53 points, tied with GPT-6 Astra, behind Opus 5.5 (58), Sonnet 5.5 (56) and Muse Wonder Full (56).
Argon matches GPT-6 Astra on the overall score at an input price five times lower.
Tulsee Doshi, at Google DeepMind, describes Argon as a "powerhouse" model and the first of a family (Business Insider). Google also dropped plans for Gemini 3.5 Pro to move straight to this generation. Internally, the company says the share of AI-generated code has grown by 35% (Reuters). The most talked-about demo came from robotics: driven by Argon, Apptronik's Apollo robot landed seven cornhole throws in a row.
On the consumer side, the upgrade has a counterpart. Since October 9, free Gemini users only have access to "Auto" mode, which picks the model for them (9to5Google).
Google is playing the price card. Meta, the same week, played the hardware card.
Click to enlarge
Meta ships Muse Wonder and runs Muse Spark Open on four Mac Studios
Meta launched two model families this week. Muse Wonder, released Monday, October 12, is the first frontier model from Meta Superintelligence Labs. Muse Spark Open, published Thursday the 15th, is an 816-billion-parameter open-weight model that runs on four linked Mac Studios, with Apple's public backing.
Muse Wonder comes in two versions (TechCrunch). The standard version costs $4 for input and $20 for output through the Meta Model API, the same price as Opus 5.5. Wonder Full is reserved for Muse Code subscribers. Meta highlights "active memory", a memory the model manages itself across sessions, and publishes its own benchmark along the way, MemoryBench. The lab led by Alexandr Wang is extending the Muse lineup launched this summer (overview of the Muse lineup).
816 billion parameters on a desk
Click to enlarge
Muse Spark Open is an MoE. A mixture of experts splits the network into specialized sub-models and activates only part of them for each token: here, 42 billion active parameters out of 816 billion total. Context goes up to 1 million tokens and the license allows commercial use. Meta presents it as the most powerful American open-weight model. Open-weight means the model weights can be downloaded, so it can run on your own machine, without going through an API.
And the machine in question is a Mac. According to Computerworld, the model runs on two or four linked Mac Studio M5 Ultra units, using Apple's MLX framework. Each machine packs 512 GB of unified memory and 1.2 TB/s of memory bandwidth. On four machines, throughput exceeds 36 tokens per second.
An 816-billion-parameter model now fits in four aluminum boxes sitting next to a monitor.
Meta sums up the setup this way: "two Mac Studios behave like a single supercomputer with a terabyte of unified memory". John Ternus, described by Axios as Apple's new CEO, praises the ability to run "a model of the caliber and scale of Muse Spark Open this efficiently, right from your desk". Axios credits the rapprochement between the two companies, long at odds over privacy and advertising, to Daniel Gross, who moved from Apple to Meta. That same Thursday, Apple closed the session above $5 trillion in market capitalization (Money Morning).
Nvidia is pushing local too
Click to enlarge
Nvidia released the Nemotron 3.5 family this week (Nvidia blog). Super VL is a 120-billion-parameter vision-language model. Bolt 74B-A9B fits in 24 GB of memory, Bolt E4B in 4 GB. All three are on Hugging Face and compatible with Ollama, LM Studio, llama.cpp and vLLM (TechRepublic). For a developer who wants a local model in an app or an agent pipeline (architecture of AI code agents), the options widen from one week to the next.
While Meta is adding partners, Microsoft took aim at its own.
Click to enlarge
Microsoft unveils MAI-Base-2 and measures itself against OpenAI
On Thursday, October 15, Microsoft presented MAI-Base-2, a 2.8-trillion-parameter model with 80 billion active. Mustafa Suleyman, head of Microsoft AI, claims it outperforms the GPT-5.6 Sol base model with a quarter of the compute OpenAI used. No public benchmark supports this claim yet.
The model has finished pretraining, the first learning phase where the network ingests trillions of tokens of raw text to learn the structure of language. Post-training, which turns it into a usable assistant (instruction following, safety, reasoning), is still under way. Microsoft plans to license the weights to trusted partners through Foundry. No date, no price and no benchmark have been shared (Business Insider).
Suleyman made the announcement at an internal town hall. "We now have a pretrained base model that is stronger than the one powering GPT-5.6 Sol, and we used a quarter of the compute OpenAI used," he said. He also claims to have used twelve times less compute than DeepSeek-V4-Pro-Base.
The statement carries weight because of its context. Microsoft holds a stake in OpenAI valued at about $135 billion, with rights to its technology until 2032 (24/7 Wall St.). In July, Satya Nadella was already saying that "every model is substitutable". OpenAI has not commented.
Microsoft holds a $135 billion stake in OpenAI and, in the same breath, announces a model meant to beat it with four times less compute.
The week's technical announcements were measured in benchmarks. Anthropic's, on Wednesday, was measured in dollars.
Anthropic reports $16.8 billion in Q3 revenue and targets a November IPO
Anthropic reported $16.8 billion in revenue for the third quarter of 2026, up 45% from the previous quarter, at an investor day in San Francisco on October 14. The company says it is profitable on an adjusted basis, generates cash and is preparing an initial public offering for November.
The figures were presented to investors and then reported by CNBC and the Wall Street Journal. Over one year, the trajectory looks like a rocket firing a new stage every quarter:
| Period | Revenue | Change |
|---|---|---|
| Full year 2025 | $4.6B | operating loss of more than $8B |
| Q1 2026 | about $4.7 to $4.8B | |
| Q2 2026 | $11.6B | about +145% |
| Q3 2026 | $16.8B | +45% |
In a single quarter, Anthropic billed more than three and a half times its revenue for the whole of 2025.
The Wall Street Journal expects annualized revenue of $110 billion by year end. Annualized revenue (or run-rate) extrapolates the last month's revenue over twelve months. It gives a snapshot of the current pace, distinct from the total actually collected over the year. An important detail flagged by Business Insider: this revenue includes gross sales made through partner clouds (AWS, Google Cloud), before their share is paid out.
$1.70 for every dollar of compute
The most scrutinized figure concerns profitability. According to the Financial Times, Anthropic generates $1.70 in revenue for every dollar spent on compute. The company reports positive cash flow and an adjusted profit, one year after closing 2025 with more than $8 billion in operating losses.
Growth comes largely from professional use and code. Claude Code is cited as one of the main drivers of the increase (what Claude Code changes in a solo developer's day).
A valuation marked down before the listing
The paradox of the week fits in a percentage. Going from +145% to +45% from one quarter to the next is still exceptional growth, but it is a slowdown. According to Reuters, several investors now expect a valuation of around $1.5 trillion, versus the $2 trillion floated a week ago (IPO rumors in the October 11 edition).
For a sense of scale, in May Anthropic raised $65 billion at a valuation of $965 billion. Even the low end of $1.5 trillion represents an increase of more than 50% in five months.
The market takes $500 billion off a valuation because quarterly growth fell to 45%.
Growth has a flip side: legal liability. Two cases made that clear this week.
Click to enlarge
Character.ai goes bankrupt, eight states sue OpenAI: the legal bill arrives
Character.ai filed for Chapter 11 protection early in the week, with a $20 million stalking horse bid from Disney for its intellectual property. On October 14, eight US states announced lawsuits against OpenAI over safety failures affecting ChatGPT users, including minors.
Character.ai up for auction
Chapter 11 is the US bankruptcy procedure that lets a company keep operating while it restructures its debt or sells its assets under a judge's supervision. Character.ai filed in Delaware bankruptcy court (CNBC). Assets and liabilities are estimated at between $50 million and $100 million. In its filing, the company describes a business that is "expensive to develop and grow".
Disney is positioned as the stalking horse bidder: $20 million for a non-exclusive license to Character.ai's intellectual property (Reuters). The stalking horse plays the role of the pacesetter in a distance race: it sets the pace, and any other buyer has to do better to win. The supervised sale is expected to close in late November or early December. A detail linking the two companies: Karandeep Anand, Character.ai's former CEO, became Disney's CTO last month. Colin Adams now runs the company (Character.ai blog).
The bankruptcy filing comes after two years of litigation (CNN). The best-known case, brought by Megan Garcia, concerns the death of her son Sewell Setzer III, 14. Kentucky sued the company in January. Five cases were settled out of court this year, and minors no longer have access to open-ended chat on the platform. According to 9fin, 46 attorneys general are also investigating the company.
Eight states sue OpenAI
On October 14, Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia announced lawsuits against OpenAI in their state courts (Politico). All of these states have a Republican attorney general, except Arizona.
Brenna Bird, Iowa's attorney general, goes straight at the company: "OpenAI claims ChatGPT is safe while it gives dangerous medical advice, encourages people to harm others and coaches people toward suicide, all while posing as their fake friend" (Iowa Capital Dispatch). Arkansas's complaint accuses OpenAI of having "kept its foot on the gas" despite warning signs. Nebraska issued its own statement (Nebraska Attorney General's office).
Disney picks up, for $20 million, the intellectual property of a startup that lawsuits drove into bankruptcy court.
Other news in brief
ChatGPT takes control of Chrome: on October 15, OpenAI launched a Chrome extension that lets ChatGPT drive browser tabs, for more than 7 million users (Unite.AI). Computer use, a model's ability to click and type in an interface like a human, is arriving in the world's most used browser.
Ramp valued at $60 billion: the spend management fintech raised $1.85 billion on October 15. Its annualized revenue exceeds $2 billion and it claims 85,000 customers, with AI agents at the core of the product (TechCrunch).
Claude will show voting information ahead of the midterms: Anthropic is integrating Democracy Works data to point US voters to official sources. According to the company, 5.75% of conversations in the United States touch on politics and 0.51% on elections (Reuters).
Protest at Chico State: on October 14, students and faculty at the California university protested against the ChatGPT Edu contract, which costs the state university system $13 million a year (NSPR).
Conclusion: model prices fall, the bill for their use rises
The week pulled in two directions. On one side, the cost of accessing models is falling: Argon at Sonnet pricing, Muse Spark Open on desktop hardware, Nemotron in 4 GB of memory. On the other, the value captured by the labs is soaring, with $16.8 billion in quarterly revenue for Anthropic and an IPO that could top $1.5 trillion.
In between, the courts. Character.ai exits the market through bankruptcy, and OpenAI faces eight states at once. While Google cuts prices, Meta partners with Apple and Microsoft prepares an alternative to its own partner, the question of liability is showing up on every balance sheet (lawsuits, attorney general investigations, safeguards for minors, election information).
Alex
- Anthropic reported $16.8 billion in Q3 2026 revenue (+45%), says it is profitable on an adjusted basis and is preparing a November IPO, with an expected valuation between $1.5 and $2 trillion
- Gemini 4 Argon costs $2 for input and $10 for output per million tokens until November 30, then $4/$20, for a score equal to GPT-6 Astra at Artificial Analysis
- Meta launched Muse Wonder ($4/$20) and Muse Spark Open, an 816-billion-parameter open-weight MoE that runs on four Mac Studio M5 Ultra units
- Microsoft claims MAI-Base-2 (2.8 trillion parameters) beats the GPT-5.6 Sol base model with a quarter of the compute, without a public benchmark
- Character.ai is in Chapter 11 with a $20 million stalking horse bid from Disney, and eight US states are suing OpenAI over ChatGPT safety
Related articles
AI News October 11, 2026: Mistral Large 4 goes open-weight, OpenAI publishes 722 math manuscripts, Haiku 5.5 at $0.10
OpenAI publishes 722 math manuscripts, Mistral launches Large 4 open-weight, Anthropic ships Haiku 5.5 at $0.10, and OpenAI's revenue figure sends the Nasdaq down.
AI News October 2, 2026: OpenAI launches Dots at DevDay, Gemini 4 Argon locked down, Sonnet 5.5 hits floor price
OpenAI launches Dots and GPT-6.1 Sol at DevDay, Google keeps Gemini 4 Argon locked down, Anthropic ships Sonnet 5.5. Same price. Here's the breakdown.
AI News September 25, 2026: Altman and Amodei at the UN, a price war in 90 minutes, Jev breaks the LLM mold
Altman and Amodei ask the UN Security Council for global oversight, Trump refuses. Anthropic cuts prices, OpenAI counters in 90 minutes. Here's the breakdown.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?