The $800 import loophole just stayed dead

Welcome back technologists🫡

One lab raised prices, one cut its prices in half, and a court kept a tax break dead, all inside the same 24 hours. Kobe slept through it; I did the math for you. (about 5 minutes)

Here's the signal today: 🛸
(ayo denk co-founder of beehiiv here’s ur intelligence report) 🕊️

👨‍💻🐕️

📦 De minimis loophole stays closed; imports cost more
🔌 DeepSeek open-sources a Claude Code rival, then raises API prices
🔌 Google halves Gemini Flash prices through 2026
🔌 OpenAI's Ultrafast mode: 14x speed, 750 tokens a second
🔌 GLM-5.3 lands with coding and cyber claims
🧮 SaaStr's account-manager math by contract size
🧠 What brains teach us about LLM memory design
🛡️ FinCEN exempts U.S. owners from ownership reports

📦 The $800 import loophole just stayed dead

The trade court said no. A federal trade court ruled August 13 that President Trump lawfully ended the de minimis exemption, which had let goods valued under $800 enter the U.S. duty-free. If you run an e-commerce store, sell hardware kits, or dropship anything physical, your landed costs are not going back down.

The plaintiff, Detroit Axle, a Michigan auto-parts distributor, sued in May 2025, arguing Trump lacked legal authority to scrap the loophole under the emergency statute he invoked. The court disagreed. Trump celebrated the decision, claiming the exemption had cost the U.S. billions of dollars in uncollected import-tax revenue. Detroit Axle's attorneys did not respond to requests for comment; an appeal could follow.

💰 Money play. Reprice any SKU still assuming duty-free imports, and check whether your suppliers' quotes quietly baked duties in. Do it before your next restock, not after the invoice arrives.

🔌 Agent Watch

DeepSeek shipped a harness. DeepSeek Harness v0.1 entered developer preview under the MIT license, an open-source alternative to Claude Code built on the premise that practically every part of the agent loop can be swapped out as a plugin. It launched alongside DeepSeek-V4-Pro, an updated flagship tuned for agentic workloads, with native OpenAI Responses API support and Codex integration. The catch: flat API pricing ends August 16 at 16:00 UTC, replaced by peak and off-peak rates, and even the discounted off-peak output prices will be substantially higher.

💰 Money play. Clone the repo and route one existing agent task through it before the new rates hit. Off-peak is now a real line item: move batch jobs off peak hours and keep interactive calls on cached input.

Google cut Flash in half. Gemini 3.7 Flash targets coding, agents, and knowledge work, and through December 31 it costs $0.75 per million input tokens and $3.75 per million output. On January 1, 2027, that doubles to $1.50 and $7.50. The launch landed three weeks after Gemini 3.6 Flash, a turnaround Google credits to developer feedback. The window is the point: you get four-plus discounted months to see whether the claimed drop in retries beats your current stack on total cost. Google gave no date for its next flagship Pro model, Reuters reported.

💰 Money play. Move one high-volume agent's inference to 3.7 Flash while the discount holds; the retry data you collect now prices your decision before the rate doubles.

OpenAI added a fast gear. Ultrafast is a new mode for GPT-5.6 Sol that OpenAI says runs at 14x standard speed, up to 750 output tokens per second, aimed at incident response, customer support, and market analysis. OpenAI's pitch: real-time speed used to mean picking a smaller model. Anthropic's Claude has a fast mode, but per TechCrunch it does not reach these speeds.

💰 Money play. If users abandon your agent mid-task over latency, switch your slowest endpoint to Ultrafast and measure completion rate, not just speed.

GLM-5.3 landed. Z.ai's new frontier coding model arrives with claims of emergent cyber capabilities and a turn at the top of the Hacker News front page on launch day. The vendor page is the spec sheet, so treat the cyber claims as vendor claims until independent evals land. Worth a spot on your eval list, not your production pipeline.

🧮 The Build

SaaStr drew the hiring line. Jason Lemkin's rule of thumb: below $8K to $10K ACV, dedicated account managers do not pay for themselves, so systematize onboarding and support instead. Above $10K, account management starts to make sense, often pooled. At $50K and up, each customer earns a named human. The real inflection sits near $2M to $3M ARR, when a first VP of Customer Success makes sense if churn or renewals hurt. One exception: near-zero CAC from pure product-led growth buys you humans earlier.

💰 Money play. Find your ACV, then count what you actually spend per account on human touch. Below $10K and hand-holding every customer? That gap is your margin leak.

🧠 Frontier

Memory, mapped. A new arXiv paper compares brains and LLMs as memory systems through shared functional questions: where information is represented, how partial cues recover associations, how new data gets written, and how memory states can be perturbed. In LLMs that maps to weights, activations, context windows, retrieval systems, and external stores. The payoff is a clean frame for a decision you make constantly: what belongs in the context window, what goes to retrieval, and what lives in a store the agent writes itself.

💰 Money play. Audit your agent's memory against the paper's four questions this week. Most solo builds overstuff the context window and underuse external stores, which is a cost problem disguised as an architecture choice.

🛡️ The Shield

FinCEN loosened BOI reporting. A final rule published August 14 in the Federal Register (FR Doc. 2026-16576) revises beneficial ownership information reporting so U.S. person owners are exempt from disclosure. If you run a U.S. LLC or corporation, a compliance chore that hung over small companies just shrank. Read the rule text before assuming your entity qualifies, since structures with foreign owners sit differently.

💰 Money play. If every owner of your entity is a U.S. person, confirm against the rule text, then cross that filing off your compliance list and tell your bookkeeper the load changed.

🎁 Level Up: the August pricing-window checklist

Two repricing events are live right now: DeepSeek's peak/off-peak switch on August 16 at 16:00 UTC, and Gemini 3.7 Flash at half price through December 31. Run this today:

Export last month's API spend by model and by hour of day.
Tag each workload: interactive (user waiting) or batch (nobody waiting).
Move batch jobs to DeepSeek off-peak hours after August 16; keep interactive calls on cached input or a Flash-tier model.
Run your top 20 real prompts through Gemini 3.7 Flash and log retries versus your current model.
Set a calendar note for December 15 to re-decide before Flash pricing doubles on January 1, 2027.
Write the new per-call cost into your pricing page math. If your margins moved more than 5%, your prices should too.

If this issue saved you money, forward it to one founder who imports anything or calls an API daily. And reply with which model survived your benchmark. I read every one.

Chase (and Kobe, who has no API costs)
Nomad Signal | nomadsignal.ai. Forward this to one builder who needs it.

See you tomorrow,

Chase & Kobe 👨‍💻🐕

One tap. Feedback goes straight to the desk.