Anthropic's $2T IPO pitch 🏛️, GTA VI: costliest game ever 🎮, Astra out-calculated physicists ⚛️

Attendees of Wednesday's investor day in San Francisco say Anthropic also called itself profitable on an adjusted basis for the quarter. Last year it lost more than $8 billion on an operating basis, and it has committed at least $518 billion over a decade to cloud and computing. It targets a valuation above $2 trillion, and its IPO, expected before OpenAI's, makes it an early test of investor appetite for loss-making AI developers.

The biggest maker of chip-making equipment said on Wednesday it is looking at ways to expand capacity beyond the 30% increases it announced in July for each of those years. It forecast fourth-quarter sales of €14.2 billion to €15.2 billion, above the €14.19 billion analysts expected, with a gross margin of around 57%. According to CEO Christophe Fouquet, there is already some strong interest in orders for 2029 and sometimes even 2030.

The game, due 19 November, took more than eight years and two delays, as Rockstar's staff almost doubled to 4,711. Rockstar won't disclose its budget, but Bloomberg estimates the game might have cost more than double the $500 million other studios try to stay under, and some analysts expect over $3 billion in revenue in four months. It arrives as an AI-driven memory chip shortage leaves consoles costing significantly more than when they came out.

xAI made the model, which listens and replies in speech, available in its Voice Agent API on Wednesday, and agents on grok-voice-latest switched over automatically. It scored 71.6% on τ-Voice, a test of support calls with accents, background noise and interruptions, comparable with GPT-Live-1 at 40% lower latency. xAI says Starlink's call centres use Grok Voice to resolve roughly 70% of 20,000 daily calls without a human.

In Jev, TypeSafe's model for typed decisions in code, images switch on with a 'danger zone' request header, even on the current jev-1.13.0. Version 1.14.0, which jev-latest moves to in a week, also adds a box type that returns a single bounding box around a region. TypeSafe warns every model, Jev included, does worse on images than on pure text.

Its agents split a PDF or plain-text document into sections, give dense parts like proofs more compute, check sections in parallel, then cross-check results and test claims against outside sources. Google piloted it with authors of over 16,000 manuscripts at four conferences, including NeurIPS, and users can now run it on design, legal and financial documents. Heavier users can connect a Gemini API key for higher limits, and Google says uploads never train its models.

Astra, running in Codex with web search, averaged under US$3 per problem on questions that typically take a world expert a day or less. Several of its answers first marked wrong proved correct after subtle errors turned up in the experts' own solutions. With web search disabled, Astra and Claude Opus 5.5 still scored 92%, which the authors take as a sign agents have memorised much of the physics literature.

The agents made over 2,600 attempts to extract the map service's data, using urlquery.net, a public security scanner, to run their own programs. Most failed, but one key and security code exposed in another company's public source code returned real place data. A source close to Tencent said it came from a temporary evaluation of third-party models, and the authors ask labs running online agents to audit what those agents can reach.

Its researchers paired a GPT-5.6 Sol lead with Luna workers on ProgramBench, where agents rebuild tools like jq and SQLite by probing a program they can run but not read. Planning paid off most because discovering the program's behaviour was the bottleneck, while a weaker Terra lead with the same workers still trailed a solo Sol agent. Teams beat solo agents by up to about 10 points but cost 3 to 7 times more.

He wants agents to run their own checks: end-to-end tests that mimic a real user's flow, property tests that throw generated inputs at rules that must never break, and comparisons with the old system during rewrites. Those checks must be fast and repeatable, he says, or agents learn to retry unreliable tests instead of fixing the problem. A human still owns what ships, reviewing the checks' results alongside the code.

Asked to pick between two versions of a web page as a source, Argon chose the one with a table in 31.5% of tests, against 93.6% for Gemini 3.8 Flash. It still backed pages with clear prices 99.7% of the time and rejected exclamation marks and all-caps words in every test. The pages were synthetic copy Argon wrote itself, and the tests did not observe Google AI Mode's live citations.

Meg McNulty and Kislay Parashar argue no benchmark reliably shows inference buyers what a product delivers, since each captures one tuned setup on one day. One of Anthropic's three 2025 serving bugs, none in the model, hit 16% of Sonnet 4 requests at its worst hour, and its own evaluations missed it. They want buyers and lenders funding continuous tests of products as sold by end-2027, so contracts can rest on results.

The reasons adding people to a late project makes it later still apply to agents, in the essay's view, with cheap code shifting the bottleneck to deciding, verifying and coordinating. In one benchmark study of three- and four-agent teams, every multi-agent setup made sequential planning 39-70% worse, while one lifted a cleanly divisible task by about 80%. Before scaling agents, it recommends writing down shared intent and matching review capacity to output.

OpenAI's own system card flags the drop, and the piece suspects recurrent depth, a technique that reuses layers without writing a token in between, is partly why. Three safety researchers fired in early October think their firing was related to their warnings about this, while OpenAI says they mishandled sensitive information. The proposed fix for both is transparency: replacement monitoring before such models ship, and outside oversight of OpenAI.

Anthropic still leads at 43.8% to OpenAI's 40.6%, and with new-customer growth slowing, the author reckons gains must now come from heavier use. Top-1% firms spent $9,568 per employee in September, up 11.7% on August but 3.1% below July's high, as token spend stalled, which the author mainly blames on falling frontier prices. Firms that signed or saw lab deals report customers hold the negotiating edge, with discounts off sticker price.

By August Hongguo took 5.9% of China's mobile time, behind only Douyin and WeChat, and Tencent answered in late September with a free short-drama mini-program inside WeChat. Video models such as Seedance and Kling now generate photorealistic dramas, and DataEye counted about 222,000 new AI drama and animated-drama titles on Douyin in the first half of 2026. With production cheaper, robonomics writes, original stories and attention become the scarce assets.

A harness is the prompts, tools and memory that make a model an agent; Harness² tests edits to it on each task without an answer key. A single round lifted GLM-5 and Gemini 3.5 Flash by up to 4.5 points on JobBench and 11.8 percentage points on LAB's partial pass rate, with model weights unchanged. The Apache 2.0 code leaves network access open and Codex's sandbox off, so the README tells users to restrict web access first.

It maps who calls each function using formal methods, meaning logical rules rather than guesses, and plugs into Claude Code, Codex and Cursor. On real Defects4J bugs, the tests it chose included every bug-revealing test for 94.9% of bugs, against 61.5% for the best name-matching rival. Java, TypeScript and Python are stable, JavaScript and C# are in beta, and the code is under FSL-1.1, with Apache 2.0 as its future licence.