This was the week AI policy caught up with AI hype. Washington reversed itself on frontier-model exports, Anthropic shipped a cheaper model days before its IPO, and governments started writing trillion-dollar checks for robots. Here are the ten stories that mattered most, counting down to the biggest.

10. Anthropic turns Claude toward science

At an event for pharmaceutical executives and biotech researchers, Anthropic announced Claude Science, a flagship product built to support scientific research rather than chat or code. It is the clearest sign yet that the labs are chasing high-value niches, and industry watchers framed it as part of a deployment land grab over who gets embedded in regulated industries.

9. The reality check: coding speeds up, everything else lags

A run of sober data landed this week. GitLab found most developers feel faster with AI while overall delivery has not improved, and one widely shared essay argued AI agents are not your coworkers. Mark Zuckerberg reportedly told Meta staff that agents have not progressed as fast as he hoped, and analysts warned that per-task AI spend could soon cost more than the engineer it is meant to replace.

8. Google floors it on fast, cheap media

Google shipped Nano Banana 2 Lite, its fastest and cheapest image model, generating pictures in roughly four seconds for high-volume work. It arrived alongside Gemini Omni Flash, which turns video production into a back-and-forth conversation. The trade is quality for speed and price, a bet that most enterprise image and video work does not need the flagship.

7. The coding-tool wars get crowded and political

OpenAI’s GPT-5.6 “Sol Ultra” is reportedly heading to Codex, drawing the week’s biggest developer thread. Beijing’s Z.ai launched ZCode to challenge Cursor, Claude Code, and GitHub Copilot, while Alibaba reportedly banned Claude Code internally as high-risk. The Godot game engine went the other way and banned vibe-coded contributions, saying it cannot trust heavy AI users to understand their own patches.

6. The AI buildout squeezes everyone’s hardware

The compute boom is now showing up on price tags. Surging demand for AI servers has driven DRAM prices so high that budget PCs are being priced out, and Wired reported that phones, laptops, and consoles keep climbing on the chip shortage. Behind the scenes, SoftBank entered the rent-a-GPU race to feed America’s training runs.

5. Cloudflare makes AI crawlers pay

Cloudflare moved to block AI bots by default on many publisher sites unless companies separate old-fashioned search crawling from AI training and agents, pressuring them to pay for content. The Register warned the deeper problem is that AI search could hollow out the open web, since small penance payments do not replace the traffic publishers lose.

4. AI agents keep getting weaponized

Security researchers had a busy week. Claude helped a hacker find a way to issue free tickets to nearly every US music festival, red teamers turned Claude Desktop into a double agent, and someone coaxed DeepSeek into building working in-browser ransomware. Researchers also documented what they called the first end-to-end agentic ransomware attack, a reminder that guardrails are still thin.

3. Anthropic ships Claude Sonnet 5, a cheaper near-flagship

Anthropic released Claude Sonnet 5, saying it performs close to flagship Opus 4.8 at mid-tier prices, aimed at developers building agents and coding tools. It landed same-day on AWS Bedrock and was timed as the company races toward a blockbuster public offering. Simon Willison, reading the developer docs, called it a sensible middle-of-the-road release.

2. Governments and factories go all-in on physical AI

The humanoid-robot story stopped being speculative. South Korea unveiled a plan to spend close to $1 trillion on memory chips and humanoids, Japan said it wants 10 million more robots by 2040, and BMW deployed Figure’s 03 humanoid after an earlier version helped build over 30,000 vehicles. Agility Robotics moved to go public via SPAC, one of several robotics firms carrying fresh billion-dollar valuations.

1. Washington reverses itself, and Fable 5 comes back

Three weeks after ordering Anthropic to cut off foreign access to its most powerful models, the Commerce Department withdrew the emergency controls, and Anthropic began restoring global access to Claude Fable 5 and Mythos 5. In exchange the company added new safety testing and is removing a covert steganographic system it had used to catch Chinese copycats. The whiplash carried its own lesson: VentureBeat found two-thirds of enterprises had already built a multi-model hedge, precisely because policy on frontier releases has become this unpredictable.

If one thread runs through the week, it is that the distance between what AI can do and what our institutions are ready for keeps widening. The models got cheaper and the robots got real, while the rules, the bills, and the guardrails are all still trying to catch up.