Trump blacklisted Anthropic from government use and the Pentagon designated it a "supply chain risk" after Dario Amodei rejected the Pentagon's ultimatum.

The crux of the issue: Anthropic wanted explicit prohibitions on mass surveillance and autonomous weapons.

OpenAI signed a Pentagon deal within hours claiming they actually got those same red lines. xAI's Grok also signed a defense deal for classified use.

The whole thing went from contract dispute to culture war really quick. 300 Google and 60 OpenAI employees signed an open letter supporting Anthropic; Claude went #1 on the App Store; Katy Perry subscribed to Claude.

Anthropic CEO Dario Amodei called OpenAI's messaging "straight up lies" in a leaked memo, which he later apologised for the tone of.

The supply chain risk designation is designed for foreign espionage threats and has never been used against a US company before. It’s wildly disproportionate for a contract dispute.

There is one excellent article you need to read that will help you understand why this happened and what implications it has for AI. Here’s a quote from it:

OpenAI released GPT-5.4 with native computer use and 1M token context. It’s also 50% cheaper than Anthropic’s flagship model Opus 4.6. Has strong benchmarks but for some reason ranked 6th on lmarena. Wouldn’t get too excited about the context window in itself; we’ve already seen how more context = worse performance in most cases. Also the pricing is a bit confusing — it increases as you fill up the context window.

OpenAI also released GPT-5.3 Instant as the new default ChatGPT model. Sam Altman is asking you to hurry and please try it soon, so they can recoup some of the 295% surge in uninstalls after their Pentagon deal.

Anthropic released Claude Sonnet 4.6 as their default Claude model. It has a 1M token context window and matches Opus 4.6 in office tasks. So it’s not just ChatGPT — Claude Free users also got a better model. I hear it’s pretty good for coding too but personally, when using Claude Code, I find myself defaulting to Opus 4.6 with maximum thinking all the time, even though I could be saving lots of tokens. Here’s how I justify that in my head: downgrading on model quality feels risky when I (as a non-engineer) can’t truly verify the quality of the code the model produces.

Google released Gemini 3.1 Pro. I saw a Google Engineer commenting that the model excels at factual retrieval and vision tasks but substantially underperforms Claude for coding despite benchmark similarity: "It is great at search, not so much action."

Google released Nano Banana 2, at half the cost of its predecessor (Nano Banana Pro). I saw levelsio, who owns a popular AI photo editing app, say his business margins went from 50% to 80% with the drop of this model. Quality-wise it’s a step back from NB Pro on some things, but it’s better on things like resemblance of people.

OpenAI launched ChatGPT for Excel in beta, embedding GPT-5.4 directly into Excel workbooks to build financial models, run scenarios, and generate formulas. It integrates financial data from sources like FactSet, Dow Jones Factiva, LSEG, Daloopa, and S&P Global. Google Sheets integration coming later. I think this can be a substantial unlock for many companies because as much as they love to talk about their advanced use of AI, we know what they actually do with their data.

Claude Cowork got a meaningful upgrade with open-source enterprise plugins and scheduled tasks. Enterprise admins can build private plugin marketplaces; scheduled tasks let Claude run workflows of your choice automatically. Claude Code also got Remote Control, where you can start a coding task in your computer’s terminal and pick it up from your phone.

Cursor released cloud agents, which run in isolated virtual machines to test UI and APIs end-to-end. This is cool. Each cloud agent runs and can do all kinds of tasks in their own little virtual machine!

Anthropic revealed competitors created 24,000+ fraudulent accounts trying to distill Claude's capabilities. DeepSeek, Moonshot AI, and MiniMax were all caught red-handed. Google has also flagged similar attempts against Gemini recently. The question is: is it okay for Anthropic et al to “distill” all the internet, books, etc but when competitors do the same back on them, it’s not?

OpenAI acqui-hired OpenClaw creator Peter Steinberger to lead its personal agents push. OpenClaw becomes an independent open-source foundation sponsored by OpenAI. Anthropic previously requested Steinberger to change the project’s name, as it was initially named Clawdbot. One company saw a legal game where the other saw a business opportunity. One thing is certain — both OAI and Anthropic are all in on autonomous agents based on their LLMs, whether the R&D to get there is done in-house or outsourced.

OpenAI landed multiyear deals with Accenture, BCG, Capgemini, and McKinsey to deploy its enterprise Frontier platform. The platform lets companies create AI coworkers with governed permissions and identities. Smart move. Getting the big consulting firms onboard means OAI is leveraging the years of trust these advisors and system integrators have built with Enterprises, making it feel like a safer choice for an executive to integrate their models.

Block cut 4,000+ employees, framing it as a proactive cut to “move faster with smaller, highly talented teams using AI to automate more work.” Their stock rose 24%. My take: this has less to do with AI and more about undoing the crazy hiring sprees done by the company right after covid when there was a zero interest rate policy. I think there will be a lot of job cuts because of AI this year, but just as often AI will merely be a convenient scapegoat.

OpenAI credits GPT-5.2 with a first original theoretical physics contribution -- a simplified gluon formula verified by Harvard and Cambridge researchers. It wasn’t GPT alone though, it was a collaborative achievement according to the paper’s author.

Goldman Sachs said AI added basically zero to US GDP in 2025, with ~90% of firms reporting zero productivity gains. This despite the massive cloud service providers having a capex nearing $500B in 2025, and predicted to hit $770B in 2026. My take: the biggest productivity gains are still at the individual level, not in orgs. With some massive exceptions, companies adopting AI have generally speaking been slow and ineffective, aiming initiatives at the wrong problems or not being ready in terms of data infrastructure (the unsexy but crucial groundwork to make it work well) or cultural change. AI is a huge economic shift and will take time to percolate through the economy.

I’m using Claude Code’s new remote control for one specific thing. Not for constantly coding on the go or at the gym — I’d rather be present where I’m at — but for those times I need to be somewhere else, but my Claude Code agent has an ongoing task. I can enable remote control, access it from my Claude app, and make sure the agent doesn’t get stuck at needing directions or permissions. In other words, I don’t try to actively develop anything from my phone, I’m merely nudging an ongoing process in the right direction so it’s finished business by the time I’m back at my computer.

I haven’t personally coded with Gemini for a while, but I’m already using Gemini Pro 3.1 actively through the API to power some retrieval workflows in the latest project I’m building.

I think Cursor’s new VM cloud agents are interesting, although I don’t have a paid Cursor subscription at the moment so not using it (I use Claude Code in a terminal inside Cursor instead). I’ve personally seen big productivity boosts from creating systems that let Claude Code run and test everything by itself (Chrome DevTools MCP, a Hetzner VPS for resource intensive tasks, and more). Bullish on tools that enable workflows like that.