Claude picked the targets

Anthropic's Claude led target selection in a Pentagon operation, while OpenAI's GPT-5.3 Instant reduces hallucinations amid a pricing war.

AI just went to war. On February 27, the Pentagon designated Anthropic a “supply chain risk to national security” and began unwinding a $200 million contract. On February 28, Claude remained active inside Palantir’s Maven system, selecting and prioritizing targets during Operation Epic Fury in Iran. We broke down the full timeline, the contract clause that broke the deal, and both sides’ positions—including Anthropic’s refusal to remove language about “analysis of bulk acquired data,” the Pentagon’s demand for unrestricted access, and Anthropic CEO Dario Amodei’s allegation that political donations influenced the outcome. Within days, 10 defense portfolio companies dropped Claude, OpenAI signed a competing Pentagon deal, and ChatGPT uninstalls surged 295% while Claude climbed to #1 on the U.S. App Store. The dispute is now headed for court—and the precedent it sets will define how AI companies negotiate ethics clauses in federal contracts.

While the military AI debate plays out, the models keep getting cheaper and the release pace keeps accelerating. OpenAI shipped GPT-5.3 Instant on March 3 with a focus on reducing hallucinations and over-cautious refusals. It delivered—hallucinations dropped 26.8% with web grounding, 19.7% without—but OpenAI’s own data shows safety refusals dropped alongside them. We looked at what the accuracy-safety trade-off looks like in practice, and the upgrade cycle isn’t slowing: GPT-5.2 Instant retires June 3, GPT-5.4 is already on the Arena leaderboard. Google isn’t waiting either—Gemini 3.1 Flash-Lite launched at $0.25 per million input tokens, one-eighth the cost of Pro, with 2.5x faster time-to-first-token. And DeepSeek V4 is preparing a trillion-parameter multimodal model built for Huawei Ascend chips—32 billion active parameters per token, a 1-million-token context window, and a projected $0.10–$0.30 per million input tokens. The pricing war is now three-sided—U.S. closed-source, U.S. open-weight, and China—and the floor keeps dropping.

The infrastructure bankrolling that race just got its price tag. Broadcom’s CEO forecast AI chip revenue “significantly” over $100 billion in 2027, backed by six anchor clients including Google, Meta, Anthropic, and OpenAI. Q1 2026 AI revenue hit $8.4 billion—doubled year-over-year—with Q2 projected at $10.7 billion and 1 gigawatt of TPUs planned for Anthropic alone this year, scaling to 3 gigawatts in 2027. Meanwhile, Swedish investigators found that Meta salespeople told customers Ray-Ban smart glasses data stayed on-device—when in fact footage was routed to servers and then to a subcontractor in Nairobi, where workers earned a few dollars an hour reviewing intimate recordings. The gap between what AI companies promise and what the data pipelines actually do is worth paying attention to.

Latest from PulseMark



The Pentagon Blacklisted Anthropic While Claude Was Selecting Targets in Iran

The contract clause, the legal challenges, and why three independent legal analyses say the designation may not hold up in court. Plus: the Palantir Maven overhaul could exceed $1 billion.

Read the investigation →



GPT-5.3 Instant Fixed the Cringe — OpenAI’s Data Shows the Cost

GPT-5.2 Instant retires June 3 and GPT-5.4 is already on the Arena leaderboard. The upgrade cycle is accelerating faster than the safety testing can keep up.

Read the analysis →



ChatGPT Uninstalls Surged 295% After OpenAI’s Pentagon Deal

The uninstall spike data, Claude’s App Store climb, and what the consumer response tells us about AI brand loyalty when military contracts enter the picture.

Read the analysis →



DeepSeek V4 Isn’t a Chip Story. It’s a CUDA Story.

A trillion parameters on Huawei Ascend and Cambricon chips—with an open-source license planned. The real disruption is proving CUDA’s lock-in can be broken.

Read the deep dive →



Meta’s Ray-Ban Glasses Filmed Bedrooms. Salespeople Lied.

Swedish regulators exposed the data pipeline: from on-device promises to Nairobi annotation centers. The supply chain of AI training data has a human-rights problem.

Read the investigation →

That’s Thursday. Seven stories, zero fluff.

— The PulseMark Team

Get the Daily Pulse

Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.

Get the Daily Pulse

Sharp AI analysis, daily. Two minutes, every morning.

Get the Daily PulseTwo minutes, every morning