One AI update just wiped out $285B

$285 billion vanished from software stocks in one trading session. Not a recession. Not a rate hike. A set of AI plugins that showed Claude doing finished legal and finance work end-to-end. The companies building AI just declared war on the companies selling software. 

Read full Story

In Today’s 5-Minute AI Digest. You will get:

1. The MOST important AI News & research
2. AI Prompt of the week
3. AI Tool of the week
4. AI Tip of the week

…all in a FREE Weekly newsletter. 

Anthropic’s new Cowork plugins for legal, finance, and marketing execute full workflows—contract review, NDA triage, compliance checks—not just text generation. Thomson Reuters fell 16%. LegalZoom dropped 20%. The market’s message: AI is competing with SaaS, not complementing it. 

Anthropic’s new flagship coordinates multiple AI instances in parallel, processes 1M tokens in beta, and launches natively inside PowerPoint—reading your slide masters and brand fonts. The play: make Claude invisible inside tools you already use. 

OpenAI’s new coding model launched the same day. 25% faster, tops SWE-Bench Pro, helped debug its own training. Also the first model rated “High capability” for cybersecurity—it chains exploits without guidance. API access delayed while safety controls catch up. 

Trump’s AI litigation task force is active, mandated to challenge state AI laws in court. Commerce Department deadline: March 11. If you deploy AI in regulated industries, the compliance landscape is about to shift. 

Opus 4.6 and GPT-5.3-Codex launched within minutes of each other. On Terminal-Bench, they’re 0.7 points apart. The benchmarks are converging—which means the benchmarks no longer decide winners. 

The real competition moved downstream. Anthropic is embedding Claude into Excel, PowerPoint, and workflows via Cowork—making AI invisible inside your existing tools. OpenAI is building Frontier—a new platform for managing AI agents as colleagues. One wants to disappear into your workflow. The other wants to replace it. Both create lock-in. Both explain why $285B in software value evaporated in a day. Read full Story

A correction: we predicted Sonnet 5 would ship alongside Opus 4.6 last week. We were wrong. Anthropic chose Office integrations over a model refresh—which tells you exactly where their priorities are. 

Prompt:
You are a senior strategist who has seen hundreds of projects fail. I’m going to share a plan. Your job: 

1. Identify the 3 most likely failure modes (be specific, not generic) 

2. For each, explain the early warning sign I’d see before it’s too late 

3. Suggest one concrete mitigation for each 

4. Name the biggest assumption I’m making that I probably haven’t questioned 

5. Rate my plan’s survival odds (%) and explain your reasoning 

Be brutally honest. Optimism is not helpful here. I need the truth I’m not telling myself. 

[Paste your plan, proposal, or strategy here] 

Swap in: a hiring plan, a product launch, a budget proposal, 

a personal goal, a move to a new city—anything with stakes. 

Why it works: LLMs default to agreement. The role assignment (“seen hundreds of projects fail”) and explicit override (“optimism is not helpful”) flip that default. The numbered structure forces specificity—no vague warnings, only concrete failure modes with early warning signs. 

Use it today: One user ran their Series A pitch through this and found a unit economics hole their CFO had missed. Works best with Opus 4.6 or GPT-5.2+ (models with strong reasoning). 

OpenAI Codex App — a coding agent you steer mid-task, like managing a junior dev 

Why you need it: Most AI coding tools are single-shot: prompt, generate, fix. Codex runs as a persistent agent—it takes multi-file tasks, sends progress updates, and lets you redirect mid-stream without losing context. Delegation, not autocomplete. 

One-liner: “Assign it a ticket. Check in later. Review the PR.” 

Rating: ⭐⭐⭐⭐ (4/5) — Powerful persistent-agent workflow, but full API access for GPT-5.3-Codex is still gated, which limits integration into automated pipelines and CI/CD. 

Key features: Mid-task steering without restart. @codex in GitHub PRs and Slack. Multi-language across CLI, IDE, web, mobile. 

Best for: Multi-file refactors, migrations, and feature builds where you want to check in periodically—not babysit every keystroke. 

codex.openai.com (requires paid ChatGPT plan) 

The tip: Opus 4.6’s adaptive thinking has four effort levels: low, medium, high (default), max. Match effort to task complexity. Translation? Low. Multi-step debugging? Max. Your bill drops without touching quality where it counts. 

Why it works: Extended thinking burns tokens on reasoning before output. On simple tasks, that adds cost and latency with zero quality gain. Lower efforts skip the deep reasoning—cutting cost and response time by up to 60%. 

Limitations: API-only for now (thinking: {type: “adaptive”}). Claude.ai handles routing automatically. Don’t use low effort on multi-step reasoning—you’ll get shallow answers on hard problems. 

Pro move: Build a routing classifier: simple lookup → low, data analysis → high, novel research → max. One early adopter reported 40% API cost reduction with no measurable quality loss on their production workloads. 

You just learned: 

The $285B selloff is the market repricing AI as a SaaS replacement—not a complement 

A pre-mortem prompt that stress-tests your plans before reality does 

Adaptive effort levels can cut API costs 40-60% on routine tasks without touching quality on hard ones 

Now implement one. 

The readers who ship one action before Friday are the ones their teams turn to when the next wave hits. 

Stay curious—and stay paranoid. 

— R. Lauritsen 

P.S. Next week: the AI tools quietly replacing $200K/year analysts at three Fortune 500 companies—and the prompting patterns they use. 

P.P.S. We called the Sonnet 5 timing wrong last week. We own it. Leaks point to codename “Fennec,” 50% cheaper, 1M context. Full breakdown the moment it drops.