The SaaS repricing, the tariff reversal, and the AI bot in your CI pipeline

Yesterday the Supreme Court struck down Trump's IEEPA tariffs 6-3, killing executive tariff authority and opening a $175-200B refund pool. The same day, Q4 GDP printed 1.4% against a 3% consensus, core PCE hit 3.0%, and the personal savings rate slid to 3.6%. Meanwhile Google shipped Gemini 3.1 Pro at 77.1% on ARC-AGI-2 — more than double its predecessor — at unchanged pricing. And a researcher published the full kill chain for a prompt-injected GitHub issue title driving Cline's Claude-based triage bot into arbitrary CI execution and npm publishing token theft.

Four stories. One day. If you run a product, a portfolio, or a platform, at least two of them touch your P&L this quarter.

The tariff reversal is a margin window, not a recovery

Roberts wrote the opinion, joined by Gorsuch, Barrett, and the three liberals. Tariffs are taxes; only Congress can tax. Six Republican senators have already gone on record against reimposition. The policy isn't coming back through this Congress.

What matters for operators: your import costs just fell, your customers are running out of money, and the Fed is boxed out of rate cuts before June. Core PCE at 3.0% and a savings rate at generational lows means demand weakens as your inputs cheapen. The winners here are importers with legal teams who file refund claims in the next 30 days and reprice supplier contracts before competitors do. Trump has already signaled he'll litigate the refund for two years — which is a feature for well-resourced filers and a wall for everyone else.

Yes, but — the executive branch is publicly attacking the ruling and slow-walking compliance. If the refund becomes a 24-month litigation slog, the balance sheet impact is smaller than the sell-side is modeling this morning. The take still holds because the forward tariff regime is what changes behavior, and that's dead either way.

The SaaS repricing has a framework now

$1 trillion in software market cap gone in three weeks. $285B off SaaS names in a single session after Anthropic's last release. Salesforce running three concurrent pricing models for Agentforce because nobody knows what replaces per-seat when the seats are AI agents.

The framework that survives the noise is brutally simple: software that does the work keeps its value. Software that documents the work is being replaced by AI that just does the work. Stripe, CrowdStrike, and Shopify are in the first bucket. DocuSign, Monday, Zendesk are in the second. Atlassian, Salesforce, HubSpot sit in the scary middle — big enough to absorb the first wave, exposed enough to face a cliff in 12-24 months.

Canva is the proof case on the other side. $4B ARR, 265M MAUs, and they publicly reframed from "design platform with AI features" to "AI platform with design tools." LLM referral traffic is growing double-digit percent. They're not fighting displacement, they're routing through it.

If you sell software, run the durability test on every product line this sprint. If your billing stack can't meter tokens, GPU hours, API calls, and outcomes as first-class dimensions, your pricing strategy is theoretical — a dashboard, not a runtime system.

The model layer commoditized in public

Gemini 3.1 Pro's 77.1% on ARC-AGI-2 is the headline. The number you should actually care about is buried in a practitioner's log of 3,177 API calls: Gemini burned 350,000 tokens to fix an Express.js bug that Opus 4.6 solved in 23,000. A 15x efficiency gap on identical outcomes. Both succeeded. One is a viable production choice at scale, the other is a margin-negative one — and which is which depends entirely on your workload.

Opus still leads on ARC-AGI-3, the interactive-reasoning benchmark that matters more for agentic work. GPT-5.2 trails at 52.9% while OpenAI closes a reported $100B+ round at an $850B valuation. The market is simultaneously pricing OpenAI's decline and funding its survival. Both can be true — but only one is a good bet.

The operator move here is the boring one: build the model-agnostic routing layer this quarter, and instrument cost-per-correct-answer on your actual production workload, not the vendor's benchmark sheet. Leadership has rotated three times in six months. Loyalty is a liability.

Your CI pipeline is now an attack surface

A prompt-injected GitHub issue title. Cline's Claude-based triage bot ingests it as instructions. The bot executes commands, poisons GitHub Actions cache, and exfiltrates VS Code Marketplace, OpenVSX, and npm publishing tokens. Compromised tokens push malicious updates to millions of developers on auto-update.

This is the pattern, not the incident. Any LLM agent that reads untrusted text and can execute commands is a remote code execution endpoint with a natural-language API. Cursor published the reference defense the same week — full autonomy inside a sandbox, human approval only at boundary crossings (network, credentials, publishing). Adopt it, or convert your LLM triage bots to read-only until you do.

Most security programs have no inventory of which AI agents hold which CI permissions. That's the audit for this week. Not next quarter.

What to do before Friday

One action per front, and none of them is aspirational. File tariff refund claims and remodel your 2026 plan against 1.5-2.0% GDP with no rate cuts. Run the "does this do work or document work" test on every SaaS product you ship or hold, and shortlist the ones that fail. Instrument tokens-per-correct-answer on your top three LLM workloads and put a model router on the roadmap. Inventory every AI bot with write access to CI, and revoke publishing tokens from anything that reads untrusted text.

The day the model got twice as smart at the same price is the day your pricing model got half as durable. Both are true. Both are Tuesday.