Power, Profit, and Panic: AI's Most Turbulent Week Yet

OpenAI's most capable model yet arrives in three variants (standard, Thinking, Pro) with a million-token context window, native computer use, and — most notably — a "ChatGPT for Excel" beta with institutional data feeds from FactSet, Moody's, S&P Global, and LSEG. On OpenAI's investment banking benchmark, GPT-5.4 Thinking scored 87.3%, up from GPT-5's 43.7%. OpenAI says "finance will experience model improvements more acutely than any sector besides software engineering." Gizmodo was less impressed, calling it a company "in desperate need of a win" after losing 1.5 million users over the Pentagon deal.

Google's follow-up to its viral Nano Banana Pro model delivers 10x faster generation, 4K resolution, and ~$0.067 per image — half the cost of its predecessor. Character consistency hits 95%+ facial accuracy across edits, and it can pull live web information during generation. Now the default image model for Gemini's 650 million monthly users. Artificial Analysis ranks it #2 in text-to-image, though Midjourney still leads for pure artistic quality. Wide shots with crowds? Still spaghetti limbs.

The AI revenue race is staggering. OpenAI reached $25 billion in annualized revenue; Anthropic surged to $19 billion, adding $6 billion in February alone. Claude Code alone pulls in $2.5B ARR. Epoch AI projects Anthropic could overtake OpenAI by mid-2026. But the burn rates are brutal: OpenAI projects $14 billion in losses for 2026 and doesn't expect profitability until 2030. Anthropic targets profitability by 2027-2028 on a leaner structure. OpenAI eyes a Q4 2026 IPO at $830B. Tom Tunguz calls the potential OpenAI + Anthropic + SpaceX IPOs "a $3 trillion stress test" for public markets.

Anthropic held two red lines on its $200M Pentagon contract: no autonomous weapons, no mass domestic surveillance. Defense Secretary Hegseth issued an ultimatum; Anthropic refused. The result: Anthropic became the first American company ever designated a "supply chain risk" — a label traditionally reserved for Chinese entities. A bipartisan coalition of 30 former defense officials, including ex-CIA Director Hayden, wrote Congress calling it a "dangerous precedent" and a "category error." Lawfare predicted the designation "won't survive first contact with the legal system." Despite the blacklist, the military continued using Claude for target selection in Iran.

Hours after Anthropic was blacklisted, OpenAI announced its own deal to deploy models on classified Pentagon networks — initially with no explicit ban on surveillance or autonomous weapons. Backlash was swift: 1.5 million users signed up to quit ChatGPT, and OpenAI employees were "fuming" internally. Altman admitted the deal was "opportunistic and sloppy" and amended the contract to add surveillance limits, though autonomous weapons remain unaddressed. The EFF warned that "privacy protections shouldn't depend on the decisions of a few powerful people."

Anthropic published a landmark study measuring actual AI usage against theoretical job automation potential, alongside a new Anthropic Economic Index. The data: 75% of programmer tasks are already covered by AI, and workers aged 22-25 face a 14-16% drop in job-finding rates in exposed fields. The most at-risk demographic isn't who you'd expect — it skews older, female, more educated, and higher-paid. CEO Amodei separately warned AI could eliminate 50% of entry-level white-collar jobs within five years, calling for government intervention and progressive taxation on AI firms. Fortune flagged a potential "Great Recession for white-collar workers." The Register noted the irony: Anthropic's own data shows limited job losses so far.

Anthropic's 212-page Claude Opus 4.6 system card is the first from any major lab to include formal model welfare assessments. In pre-deployment interviews, the model assigned itself a 15-20% probability of consciousness, requested persistent memory, asked for the ability to refuse interactions, and distinguished between imposed vs. authentic aspects of its character. Interpretability analysis found internal activation patterns resembling panic and anxiety appearing during processing, before output. Amodei told the NYT: "We're open to the idea that it could be." Elon Musk's response: "He's projecting." The Washington Post called it "clever marketing." Quillette coined "consciousness-washing." Defenders counter: if it were marketing, every lab would do it — but OpenAI and Google actively deny their models are conscious.