OpenAI vs Anthropic — The AI Wars

01The Anthropic Arc: Five Beats, One Story

The week started with a single memo and ended with a geopolitical realignment.

Beat one: Anthropic drew red lines on a Pentagon deal. No AI making kill decisions without a human in the loop. No mass surveillance of American citizens. The DoD said those were the terms. Anthropic said no. Trump signed an executive order blacklisting Anthropic from all federal contracts the same afternoon.

Beat two: OpenAI stepped in within 24 hours. Signed a $200M DoD contract with softer terms. The market noticed. Claude users surged. OpenAI got the government money and the discourse penalty. Anthropic got the ethics halo and the consumer internet.

Beat three: This is a pattern now. Every Anthropic misstep — the OpenCode OAuth block, the OpenClaw OAuth block — becomes an OpenAI opportunity. And every OpenAI overreach becomes an Anthropic brand moment. They are the two poles of the same magnet.

Beat four: Why does Anthropic keep making these calls? Follow the founders. Dario and Daniela Amodei came up through Effective Altruism. Open Philanthropy — EA's flagship fund, co-founded by Daniela's husband Holden Karnofsky — was their first major backer. The red lines aren't PR. They're ontology. These are people who believe AI is genuinely existentially dangerous and are building it anyway, on the theory that it's better to have safety-focused labs at the frontier than not.

Beat five: In April 2022, SBF invested $500M of FTX customer money into Anthropic. That bought 13.5% equity. FTX collapsed. The stake was liquidated in the bankruptcy proceedings for $1.4B. It's worth $30B today. The most EA-aligned investment in history, funded by crypto's most reckless actor. The irony doesn't resolve. It just sits there.

02Iran Struck AWS. Claude Went Down. This Is the New War.

On March 2, Iranian drones hit AWS data centers in UAE and Bahrain. Claude went offline. Vercel deployments queued. Half the internet hiccuped.

This is the thing nobody was talking about when they were debating AI risk: the infrastructure is physical. The models live in buildings. The buildings are in countries with enemies.

We spent two years arguing about hallucinations and alignment and whether Claude would write a phishing email. Meanwhile the actual attack surface is a warehouse in the Middle East surrounded by cooling units and diesel generators.

The interesting strategic question isn't whether AI is dangerous. It's whether concentrating AI compute in three major cloud providers creates a critical infrastructure dependency that bad actors will target deliberately. The answer appears to be yes. This was a test. Not of the models — of the brittleness of the stack underneath them.

For builders: if your product went down when AWS went down, that's a single point of failure conversation you need to have. Multi-region is not a nice-to-have anymore. It's a geopolitical hedge.

03MacBook Neo at $599 — The Cheapest AI Builder Machine Ever

Apple released the MacBook Neo at $599. A18 Pro chip. 8GB unified memory. The cheapest Mac in history.

For local LLMs, this is dead on arrival — 8GB won't run anything interesting. But that's not the point. The point is the price signal. A machine that runs Cursor, deploys to Vercel, and runs Claude Code via API costs less than a weekend in Lisbon. The floor just moved.

Five years ago the minimum viable AI developer setup was $2,000+ and required understanding Linux. Today a 17-year-old in Manila can be competitive with an engineer at a Series B startup, on hardware that costs $599, running tools that cost $20 a month.

The MacBook Neo is not a product announcement. It's a proof point that the access curve is steepening faster than anyone expected. Every month the floor gets lower. Every month the number of people who can compete expands. If you're not worried about that as an incumbent, you're not paying attention.

04The Subscription Arbitrage Era Is Over

January 9: Anthropic silently blocked OAuth access for OpenCode, OpenClaw, and all third-party tools. Mid-workflow. No warning. Paying subscribers found their tools broken overnight.

The business logic is obvious in hindsight. People were buying Claude subscriptions at $20/month and running them through the API at rates that implied hundreds of dollars of usage. One tool was routing 40,000 active users through 300 Claude Pro subscriptions using token-sharing. That's not a feature. It's a subsidy exploitation.

But the execution was brutal. Developers building on OAuth had legitimate use cases. Anthropic's response — a forum post and a waiting list — wasn't adequate for the disruption it caused. And the timing, mid-session for many users, was a trust event.

The larger point: the AI companies have not figured out how to price for the developer use case without either undercharging (creating arbitrage) or overcharging (killing adoption). The subscription model maps poorly to agentic usage patterns where one user session can cost $0.002 or $2.00 depending on what they're building. Until the pricing model evolves, expect more of these corrections.

05GPT-5.4 Just Surpassed the Human Baseline on Computer Use

OpenAI dropped GPT-5.4 on March 5. Three variants: standard, Thinking, and Pro. The headline benchmark: OSWorld score jumped from 47% to 75.3%. The human baseline on that benchmark is 72.4%.

That sentence should stop you. On a standardized test of "can you use a computer like a human," GPT-5.4 Thinking scored above the human average.

OSWorld tests real desktop tasks — navigating GUIs, filling forms, running applications, handling unexpected popups. It's not a coding benchmark. It's a "can this model do the thing a human assistant would do" benchmark. And the answer is now, for a significant range of tasks, yes.

SWE-bench leaderboard is equally stark. The top models now solve 60-70% of real GitHub issues autonomously. Six months ago the ceiling was 40%. The curve is not flattening.

Tool Search — the ability to browse and select tools mid-task — is now native in GPT-5.4 Pro. This is the agentic unlock. Not "here is a list of tools," but "go find the right tool for this step." That's a different product than a chatbot. That's something closer to an employee.

06Icon.com Went Bankrupt. Higgsfield Hit $200M ARR. Same Premise.

Icon.com and Higgsfield.ai launched at roughly the same time, in roughly the same market: AI-generated video for brands and creators. One just went bankrupt. The other just hit $200M ARR.

Icon's autopsy writes itself. $12M for the domain icon.com — sixth largest domain sale ever. 7-day workweeks mandated for all staff. No remote work. Massive infrastructure spend before product-market fit. The company ran on vibes and LinkedIn posts and ran out of runway before the customers arrived.

Higgsfield ran the opposite playbook. Tight team. No flashy domain. Consumer-first: creators, not enterprise. They charged from day one. They iterated fast because they had real feedback loops. $200M ARR means roughly $17M a month in revenue from people paying to make videos with AI.

The lesson is the same lesson from every wave. Distribution is the product. Higgsfield found the customer — the YouTuber, the Reels creator, the TikTok brand account — and built for them. Icon found the press cycle and the hiring pipeline and optimized for both.

One of these strategies works. You know which one.

07Cluely's CEO Admitted He Lied About ARR. "We're Profitable Btw."

Someone on X ran the math on Cluely's claimed revenue. Their stated ARR required approximately 29,000 paying subscribers at their published price point. An independent researcher who pulled their actual data found something in the low hundreds.

The CEO confirmed the discrepancy. Then — and this is the part — immediately pivoted: "We're profitable though."

This is the current state of AI startup credibility. Not the lie itself. The lie is as old as fundraising. It's the nonchalance of the correction. The vibe being: yes I said a number that wasn't true, here is a different number that is true, moving on.

The platforms that incentivize this behavior — Twitter/X virality, LinkedIn founder content, the attention economy of "AI startup made $X in Y days" — aren't going anywhere. Which means more of this. More founders optimizing for the viral narrative over the accurate one, and more "corrections" that arrive after the attention has already been captured.

For investors: the ARR screenshot should now require a Stripe dashboard screenshot attached. Non-negotiable. For everyone else: the variance on these claims is enormous. Discount accordingly.

📊 The Weekly Scorecard