Anthropic wants to make drugs now

PRODUCTS

Chinese AI lab Z.ai released a free coding tool this week, and its underlying model is close enough to Claude and GPT-5.5 that a lot of developers are going to notice.

The tool is called ZCode, a desktop coding environment built around Z.ai's newly released GLM-5.2 model. It's aimed squarely at developers currently paying $20 to $200 a month for Cursor, Claude Code, or GitHub Copilot. The pitch is simple: free access to a coding agent that gets within striking distance of the paid tools.

And the numbers actually hold up. On Z.ai's Hugging Face model card, GLM-5.2 scores within a point of the best model in the world on FrontierSWE, the benchmark for complex coding tasks:

+ GLM-5.2: 74.4

+ Claude Opus 4.8: 75.1

+ GPT-5.5: 72.6

The picture isn't uniform — on harder long-horizon benchmarks, GLM-5.2 still trails Opus by half. But for the kind of coding most developers actually do day to day, it's competitive. As David Sacks put it on the All-In podcast, American companies can't afford to slow down when open Chinese models are catching up this fast.

Then there's the pricing:

+ ZCode is free to use with GLM-5.2, at least through September.

+ API access costs a fraction of what Anthropic or OpenAI charge for their frontier models.

+ Using GLM-5.2 inside ZCode gives developers 1.5x the effective quota versus API access alone.

That isn't subtle. Z.ai is betting that heavily subsidized access to a nearly frontier model will pull developers away from paid Western tools long enough to build new habits. And these aren't half-baked open-source models that need fine-tuning to be usable anymore — they're plug-and-play, out-of-the-box tools that just work.

None of this is happening in a vacuum. Just weeks ago, we covered how Anthropic was quietly fingerprinting Claude Code users to defend against distillation attempts from Chinese labs. Z.ai was on that list. Now one of the labs Anthropic was trying to keep out has shipped a serious, free alternative to the product itself.

The big question is whether enterprise buyers will care. Regulated industries in the US and EU will likely refuse Chinese models in their stack regardless of price or performance — that's a real ceiling. But independent developers, startups, and everyone building outside of banks and hospitals tend to care about what works and what's cheap. On both counts, ZCode delivers.

The pricing conversation around AI coding tools has been getting uglier for months, with Cursor and Claude Code both charging more while developers argue about whether any of it is worth the money. Z.ai's response is to remove the argument entirely by giving away a tool that's roughly good enough. That won't kill Cursor or Claude Code, since enterprise contracts and best-in-class performance still matter for the companies that can afford them. But it does put a ceiling on what those tools can charge going forward, because the moment a free option gets within a percentage point of the best paid one, the math changes for everyone downstream. Anthropic can keep its enterprise customers. What it can't keep is the assumption that coding agents are inherently a premium product.

BIG TECH

Selling AI software turned out to be the easy part. Getting it to actually work inside a company is where the money and the headaches live — and Microsoft just made that its next big bet.

On Friday, Microsoft announced a new subsidiary called Frontier Company, backed by $2.5 billion and 6,000 engineers whose entire job is to embed inside customer companies and make AI deployments actually pay off. The model isn't new — Palantir built its business sending elite engineers to wire software into whatever weird workflow a customer actually runs. Now every major cloud company is copying that playbook because a Copilot license, on its own, isn't landing.

The scale of the bets tells the story:

+ Microsoft: $2.5 billion, 6,000 embedded engineers, launched as a full subsidiary.

+ AWS: $1 billion for its own forward-deployed engineer org focused on agentic AI.

+ Palantir: Q1 2026 U.S. revenue growth of 104% year over year, quietly proving the model works.

The reason everyone's moving at once: AI pilots keep dying in production. A survey from Orgvue found that 92% of organizations have invested in AI, but 78% say those projects have either failed or stalled — and 83% plan to spend more this year anyway. The model is usually fine. The hard part is everything around it: the data, the processes, and the people who need to trust it enough to change how they work. Even Zuckerberg has been telling Meta employees that AI agents haven't progressed as fast as he'd hoped.

For customers, having engineers from Microsoft or AWS camped inside your building is a mixed blessing. The NFL's CIO said AWS's embedded engineers helped launch products like NFL Fantasy AI in weeks instead of months. That's the sales pitch. The quieter reality is that every hour those engineers spend in your stack is an hour spent wiring your future to their models and their infrastructure. Palantir customers already know this. Microsoft's are about to.

The industry spent two years arguing about which model was best, and the answer turns out to matter less than which vendor will send a team of humans to your office to make the thing work. Microsoft and AWS are quietly conceding that software alone can't close the gap between what AI does in a demo and what it does at your company, so they're spending billions to hire their way across it. The winners of the next phase of AI are probably going to be the companies with the most engineers willing to show up in person and stay until something actually ships. Benchmark leaderboards are becoming a sideshow to what happens on the customer's laptop at 4 p.m. on a Tuesday.

BIG TECH

The safety-first AI company is trying to become a drug company. That's a much harder pivot than it sounds.

On Tuesday, Anthropic launched Claude Science, a version of Claude wired into tools scientists actually use — Benchling, PubMed, 10x Genomics, BioRender. But the bigger news came in the same breath: Anthropic is now running its own internal drug discovery programs, targeting diseases Big Pharma has largely written off as unprofitable, according to the company.

The head of Anthropic's life sciences work told STAT the goal is to change life sciences the way Claude Code changed programming — and that Anthropic wants to "live it" alongside the researchers it's selling to. In other words, the company is putting its own models to work on actual drug candidates instead of just licensing the tech and walking away.

This didn't come out of nowhere. A few months ago, Anthropic paid around $400 million to acquire Coefficient Bio, a team already building agent systems for biology. And John Jumper, the Nobel-winning AlphaFold researcher, quietly joined Anthropic from DeepMind earlier this summer. The talent — and the money — are moving.

The problem is what everyone in this field keeps running into:

+ Designing molecules is the easy part. A recent Latent Labs paper put it plainly: the bottleneck isn't AI-powered design, it's lab testing. One round of experiments costs $80,000 to over $1 million.

+ No AI-generated drug has actually reached the market yet. Every AI drug company you've heard of has raised enormous rounds and delivered mostly slide decks.

+ Even believers are skeptical. The CEO of Novartis — one of Anthropic's biggest customers — told a Harvard audience last year he had a "somewhat muted view" of AI's impact on drug discovery.

Anthropic already has real enterprise traction here. Sanofi, Genmab, and Novo Nordisk use Claude for drug development workflows, and Novo Nordisk has called it the standard for how it automates content. But that's mostly document work. Turning that into an approved drug is a completely different game — one that has humbled every AI-first company that's tried it.

Anthropic has done the thing every AI company eventually has to do, which is pick a hard problem in the physical world and try to solve it. Drug discovery is a strange choice, because it's the one field where the last decade of AI hype has produced the fewest actual products. Every AI drug company you've heard of has raised enormous rounds, promised revolutions, and delivered mostly slide decks. What makes Anthropic worth watching is that it's not selling to pharma from the outside, it's trying to become pharma from the inside, with the model quality and the capital to actually take the swing. Whether that ends in a real drug or another expensive lesson about how biology doesn't care how smart your chatbot is, we'll find out in about a decade.

IN OTHER NEWS

+ Alibaba bans Anthropic's Claude Code after employees discover an alleged hidden China-detection backdoor

+ An AI agent called JADEPUFFER ran a full ransomware attack on its own , fixing its own coding error in 31 seconds

+ Zuckerberg admits Meta 'miscalculated' its AI overhaul after cutting 8,000 jobs

+ OpenAI is in talks to give the U.S. government a 5% equity stake worth $42.6 billion

+ Tesla expands Robotaxi to Miami , marking its third state for driverless rides

+ GM installed 50 robots at its Detroit factory and idled over 1,000 workers — the UAW is furious

+ Samsung unveils a $44 billion plan to build a Physical AI cluster for humanoid robots and next-gen batteries

+ A leaked video reveals Project Aion, Microsoft's prototype for an AI-first version of Windows built entirely around Copilot

WHO'S HIRING IN AI

+ Scale AI — Head of Global Policy

+ Perplexity — Internal Communications Manager

+ Airbnb — Senior Manager, Product & AI Policy

+ Google — Senior AI Strategist

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ Google Maps: A new "Ask Maps" feature lets you type complex questions like "coffee shops with charging stations and short lines" and get conversational, personalized answers with a custom map

+ Adobe Firefly: Google's Gemini Omni Flash video model just joined Firefly's partner lineup, letting creators generate and edit video with AI directly inside Adobe's creative suite

+ Meta Pocket: Meta quietly launched an app where you type a sentence and it builds a playable mini-game you can share, remix, and discover in a TikTok-style feed

+ Glaze by Raycast: Describe the app you want in plain English and it builds a real native Mac desktop app — icon in your dock, works offline, with full access to local files and APIs

+ Google Drive: Ask Gemini in Drive now works on mobile, so you can ask questions about your files from your phone instead of opening documents one by one