AI Wednesday: Claude Opus 4.7 Ships with 2M Context, Microsoft Recall Goes Global, and the Week in AI Review
Anthropic launches Claude Opus 4.7 with 2 million token context window. Microsoft Recall finally ships worldwide. Plus the biggest AI stories this week.
Massive week. Two stories completely swallowed everything else, and honestly? Each one deserves a full standalone article. But we are doing the roundup, so here we go.
Claude Opus 4.7: 2 Million Tokens of Context
Anthropic dropped Claude Opus 4.7 on Monday. The headline number: a 2 million token context window. Roughly 1.5 million words. An entire codebase. Hundreds of pages of legal docs. Every research paper on a topic. All at once.
I fed it a 400-page technical manual yesterday. The model pulled details from page 12 when answering questions about page 380. Previous versions lost the thread around 200K tokens. Completely different game now.
Benchmarks tell the same story. Opus 4.7 hit new highs on MMLU, HumanEval, and a few reasoning tests I track closely. The coding gains show up in real work too, not just numbers on a chart. Multi-file refactors feel natural. Earlier models hesitated where this one just does it.
Price tag is steep though. $15 per million input tokens, $75 per million output. Most hobby projects cannot justify that. But enterprise workflows where you replace hours of manual document review? The numbers pencil out quickly.
If you have been following our coverage of Claude's evolution, this one hits different. Context size was the wall holding back serious AI use. Anthropic just smashed through it.
Microsoft Recall Ships Worldwide
Over a year of controversy, delays, and privacy rewrites later, Microsoft Recall is finally here. Worldwide. Quick recap for anyone who missed it: Recall takes constant screenshots of your PC, uses AI to index all of it, and lets you search anything you have ever seen on screen. Handy. Also kind of creepy.
Microsoft clearly heard the screaming. The version shipping this week has protections that did not exist in the original demo:
- Local processing only. Screenshot analysis stays on-device. Nothing leaves your machine.
- Biometric gate. Windows Hello required before you can touch Recall history.
- Sensitive content filtering. Automatically skips incognito windows, password fields, and flagged content types.
- Per-app opt-out. Block specific apps from being captured.
Is it enough? Privacy folks remain skeptical. Fair. A machine watching your screen 24/7 crosses a line for a lot of people, no matter where the data sits. But Microsoft learned something from the initial launch disaster. The controls are solid this time. Real ones.
Ships first on Copilot+ PCs with dedicated NPUs. Wider availability coming later this year.
The Week in Review
Those two stories stole the show, but plenty more happened. Stuff that caught my eye.
AlphaFold 4 tackles protein interactions. DeepMind's new version goes beyond predicting individual protein structures. Now it models how proteins interact with each other and with drug molecules. Huge deal for drug discovery. Pharma companies will be making noise about this. Loud noise.
China's AI agent regulations kick in. Starting this week, any AI agent running in China must register with regulators and spell out its autonomous capabilities. Covers everything from customer service bots to research agents. We touched on similar moves earlier this week.
OpenAI IPO gets more concrete. Reuters says OpenAI is eyeing early 2027 for going public. Projected valuation over $300 billion. The company is reshuffling its nonprofit board to get ready. Some people are not thrilled.
Apple hints at WWDC AI moves. Apple confirmed big AI announcements for its June event across iOS, macOS, and a new developer framework. Details are scarce. Rumors point to on-device models that compete with current cloud stuff. We are paying attention.
DeepSeek V4.1 specs leak. A credible leak hints that DeepSeek's next model breaks past 1 trillion parameters using a mixture-of-architecture approach. Chinese AI labs keep closing the gap. Fast. We flagged some of these patterns in Monday's roundup.
What I Am Watching Next
2 million tokens of context changes the math on what AI tools can do. I am putting together a hands-on guide for using Claude Opus 4.7 with big codebases. Should land next week.
Microsoft Recall will be the most argued-about AI feature of the year. I want a full month of testing before I call it. First impressions? Mixed. But interesting.
FAQ
Q: Is Claude Opus 4.7 worth the price for individual developers?
A: For everyday tasks, no. Sonnet 4 handles normal coding and writing just fine and costs way less. Opus 4.7 earns its price when you actually need to chew through huge amounts of text or code in a single session. Codebase analysis. Legal review. Research synthesis. That kind of work.
Q: Can I turn off Microsoft Recall completely?
A: Yes. Flip it off in Windows Settings under Privacy and Security. You can also keep Recall installed but kill screenshot capture while keeping the search UI around. Microsoft made the opt-out simple this time. No hoops.
Q: How does the 2 million token context stack up against Google's Gemini models?
A: Gemini 1.5 Pro also does 2 million tokens, so the window size matches now. The real question: which model actually uses that context better? Early side-by-side tests show Claude Opus 4.7 ahead on long-document reasoning. Gemini holds its own on multimodal inputs inside big contexts. Competition here is healthy.
Q: When do Apple's AI announcements actually ship?
A: WWDC hits in June, but Apple usually announces stuff that ships in the fall with new hardware. If the pattern holds, developer betas in July, public releases September or October. Nothing locked in yet.