AI News: April 27, 2026

1. Mill Valley Home Listed for Anthropic Equity Marks Private-Stock Real-Estate Crossover

Anthropic. A 13-acre Mill Valley property is being offered exclusively in exchange for Anthropic equity, with the seller — investment banker Storm Duncan, who paid $4.75M for the home in 2019 — pitching it as a portfolio rebalance away from real estate and toward AI exposure. The structure lets the buyer keep 20% of the upside on the shares exchanged for the duration of the lockup period, an unusually explicit attempt to bridge illiquid private AI stock into a tangible asset class. The listing is the latest signal that secondary demand for shares in top-tier private AI labs has hardened enough to function as a near-currency for high-value transactions in the Bay Area, even before any IPO. Source

2. Epoch / Ipsos Survey: 80% of Weekly Claude Users in the US Earn Over $100K

Anthropic. A joint Epoch AI / Ipsos US survey finds that 80% of weekly Claude users live in households earning more than $100,000, well above Microsoft Copilot (64%), ChatGPT and Grok (56% each), Gemini (56%), and Meta AI (37%). The relative skew is large, but Claude’s absolute reach is small: among high earners, ChatGPT leads at 37% weekly use versus 6% for Claude. Read alongside Anthropic’s own Project Deal research showing stronger agents extract systematically better terms in negotiation, the demographic data sharpens a real concern — if higher-capability AI assistants are concentrated among already-wealthy users, agent-mediated commerce could quietly entrench existing economic gaps rather than narrow them. Source

3. “Semi-Executable Stack” Model Reframes the Surface Area of AI Engineering

Chalmers / Volvo Group. Researchers from Chalmers University and Volvo Group published a model treating AI-augmented software engineering as six concentric rings — code at the center, then prompts, workflows, guardrails, organizational decision routines, and regulatory compliance at the outer edge — and argue that the scarce engineering skill is judging which ring a change actually belongs in. They observe that nearly all current AI engineering research targets the innermost ring (code generation and code agents), leaving the outer rings without comparable testing, monitoring, or accountability discipline, and warn that organizations treating AI as a pure efficiency play will miss the system redesign work the technology demands. The framing is a useful counterweight to “AI replaces engineers” narratives: it argues that the headcount disappearing at the inner ring needs to reappear with different skills — architectural judgment, governance, and institutional fit — at the outer ones, and that reliability concerns at the boundary are engineering problems, not philosophical ones. Source

4. BankerToolBench Reveals Where AI Agents Break in Real Knowledge-Work Pipelines

Handshake AI / McGill. A new evaluation built with nearly 500 working investment bankers tested nine frontier models on the actual artifacts a junior banker produces — Excel models with live formulas, pitch decks, and research memos — graded against rubrics averaging 150 criteria per task. None of the model outputs cleared the bar for direct client delivery: GPT-5.4 led with 16% rated as acceptable starting points, while Claude Opus 4.6 produced visually polished spreadsheets that hardcoded values instead of formulas, breaking scenario analysis. The four most common GPT-5.4 failure modes were code bugs (41%), broken business logic (27%), aborted runs (18%), and fabricated data (13%) — a useful taxonomy for any team designing tool-using agent systems in regulated knowledge-work domains, where structural correctness (formulas, references, citations) matters more than surface fluency. Source

5. Sam Altman Publishes “Our Principles” Framing OpenAI’s AGI Stance

OpenAI. Sam Altman published a short essay laying out five principles meant to guide OpenAI’s work through the AGI transition, emphasizing user empowerment, decentralized access to capability, and a deliberate framing of superintelligence as something whose benefits should accrue to “billions of people” rather than a small set of incumbents. The piece reads as a public-facing values statement aimed as much at policymakers and prospective enterprise buyers as at researchers, anchoring OpenAI’s positioning while regulatory debates around AI concentration intensify in the US, EU, and UAE. There is no new product or capability attached, but the document is likely to be cited in OpenAI’s policy submissions and partnership discussions in the coming weeks. Source

6. OpenAI Folds Codex Back Into the Main Line With GPT-5.5

OpenAI. OpenAI has retired the standalone Codex coding model line, with Head of Developer Experience Romain Huet confirming that GPT-5.3 was the last dedicated Codex release and that GPT-5.5 absorbs all coding-specific capability into the base model. The merged model is described as substantially better at agentic coding and computer use, using fewer tokens on the same coding benchmarks while charging roughly 20% more per task on net. This is the second time OpenAI has discontinued the Codex brand — the original was killed in 2023 and revived in May 2025 — and the move reflects a broader product simplification toward a single frontier line that handles both general reasoning and code-heavy agent loops. Source

7. OpenAI Tells GPT-5.5 Developers to Throw Out Their Old Prompts

OpenAI. OpenAI released a GPT-5.5 prompting guide arguing that prompts tuned for GPT-5.2 or 5.4 actively degrade performance on the new model, and recommending developers start from a minimal, outcome-focused baseline. The guide pushes a seven-part structure — role, personality, goal, success criteria, constraints, output format, stop rules — and tells teams to default to “low” and “medium” reasoning effort before reaching for higher settings, reserve absolute language (“ALWAYS”/“NEVER”) for genuine invariants like security policy, and explicitly set retrieval budgets for fact-based answers. OpenAI suggests using Codex (the agent, not the model) to bulk-rewrite legacy prompts, an unusually direct admission that GPT-5.5’s behavior shifts are large enough to require systematic prompt migration rather than incremental tweaks. Source