The Signal — March 18, 2026

OpenAI released two new models yesterday. Anthropic launched a think tank. And Sears accidentally exposed 3.7 million customer conversations to the open internet.

OpenAI Ships GPT-5.4 Mini and Nano

Two weeks after releasing GPT-5.4, OpenAI dropped smaller versions: GPT-5.4 mini and GPT-5.4 nano.

The mini model runs at $0.75 per million input tokens and $4.50 per million output, with a 400k context window. That's roughly 3x more expensive on input than GPT-5 mini, but OpenAI says it approaches GPT-5.4's performance on coding benchmarks like SWE-Bench Pro and OSWorld-Verified while running more than twice as fast. Free and Go-tier ChatGPT users get access to mini immediately.

Nano is API-only at $0.20/$1.25 per million tokens. Simon Willison ran the numbers: you could describe 76,000 photos for about $52 with nano.

The pitch is the "subagent pattern," where a larger model handles planning while mini or nano run parallel subtasks like searching codebases or processing documents. The Decoder noted the price hikes are steep compared to predecessors, up to 4x on nano inputs. Whether the capability gains justify the cost depends entirely on your workload.

Sources: OpenAI Blog · ZDNet · Simon Willison · The Decoder

Anthropic Launches the Anthropic Institute

While still locked in a legal fight with the Pentagon over its supply-chain-risk designation (a story we've been covering extensively, most recently in The Strategic Tech Cycle), Anthropic announced a new research institution on March 11: the Anthropic Institute. We missed it last week, but it deserves coverage.

Co-founder Jack Clark will lead it with a new title: Head of Public Benefit. He previously served as Head of Public Policy. The institute merges three existing Anthropic research teams to study AI's effects on the economy, labor markets, safety, and societal values. Sarah Heck, formerly Head of External Affairs, takes over the public policy team.

The timing is pointed. Launching an independent-ish research body focused on AI's societal impact while the Defense Department is actively trying to replace you sends a specific message: Anthropic wants to be seen as the company that takes the long-term consequences seriously, even at a cost.

Whether that positioning holds up depends on what the Institute actually produces. A think tank staffed by the company it studies has obvious credibility questions. But the alternative, frontier labs funding no external research into their own effects, is worse.

Sources: Anthropic · The Verge · CIO.com

Sears Chatbot Leaked 3.7 Million Customer Conversations

Security researcher Jeremiah Fowler found three publicly exposed databases belonging to Sears Home Services. They contained 3.7 million records: chat logs, audio recordings, and text transcriptions from phone calls with the company's AI assistant, "Samantha." The data spans 2024 through 2026.

The exposed conversations included personal details customers shared while scheduling appliance repairs. Names, addresses, service histories. Anyone who knew where to look could access them without authentication.

Sears still operates five physical stores and an appliance repair service. The AI chatbot was a modernization effort. What it modernized, apparently, was the speed at which customer data could be made public.

This is the kind of breach that gets buried under the AI hype cycle. No frontier model, no billion-dollar valuation. Just a company that deployed an AI customer service tool without securing its data stores. The mundane version of AI risk is often the one that actually hits people.

Sources: Wired · DataBreaches.net · ExpressVPN Blog

On the Editor's Desk

A report surfaced claiming Anduril won a $20 billion Army contract for AI-driven defense systems. We're holding it. The sole source is "The Defense News," not the established Defense News at defensenews.com. Different outlet, unknown credibility. A $20B contract would be too big for every major outlet to miss.

The pipeline also flagged companies openly signaling they're replacing workers with AI. Block cut about 40% of staff in February under Jack Dorsey's directive to stop hiring for roles AI can fill. HP announced 6,000 job cuts through 2028 while ramping AI investment. Both are real, but neither is fresh. The pattern is worth tracking even if the individual data points aren't new.

Mistral released Small 4 on March 16: 128 expert modules, 119 billion parameters, 256k context window. It landed in our pipeline at significance 2 and didn't get full review. Worth noting.

Our pipeline's scoring model continues to overvalue SEO-optimized blog titles and undervalue government contracts. Today it scored a cookie consent marketing post at significance 6 and a potential $20B defense contract at significance 1. Recalibration is ongoing.

Corrections

The original version of this edition stated GPT-5.4 mini and nano arrived "two days" after GPT-5.4. The actual gap was approximately two weeks (GPT-5.4 released March 5, mini/nano on March 17). The Anthropic Institute section was also originally framed as breaking news; the announcement was made on March 11, a week before this edition. Both errors have been corrected above. We regret the mistakes.