The 20x cost cut that could become a liability

A San Diego developer named Stu Clott reportedly used to pay about $10 for an hourlong coding session with Claude. He switched to DeepSeek. The same work now costs him less than 50 cents. That is reportedly a 20x price cut for tasks he says he cannot tell apart.

This is not a fringe hack. US companies are pulling Chinese models like DeepSeek and Z.ai into real workflows as frontier US token prices climb. Rest of World reported this trend on June 17, 2026. The reason is brutal math: NavyaAI found per-token prices fell roughly 99.7% since the GPT-3 era, yet enterprise AI bills tripled over the same window.

Here is the part nobody planned for. Uber burned through its entire 2026 AI coding budget by April, according to reporting cited by Forbes on July 2, 2026. And I think the enterprises copying that spend pattern are about to make the biggest procurement decision of the decade. Here is how to think about it.

The Three-Tier Reckoning

Cheap tokens did not make AI cheap. That is the whole trap.

Agentic workflows multiply token usage 50 to 500 times per task, per NavyaAI. So even as unit prices collapsed more than 90%, total spend exploded. NavyaAI also found 72% of production AI cost sits outside the model invoice, in orchestration, retries, and retrieval. The model bill is the part you see. The rest hides.

Call the response the Three-Tier Reckoning. Enterprises are learning to sort every task into one of three buckets by how much brainpower it actually needs.

Tier 1 is high-volume, low-stakes work: summarization, ticket triage, classification. Tier 2 is mid-capability reasoning. Tier 3 is frontier work where a wrong answer is expensive. The reckoning is simple to say and hard to do: stop paying Tier 3 prices for Tier 1 work.

That is the leak. Chinese models like DeepSeek are flooding into Tier 1, where "good enough" wins on price alone.

Why the Cheapest Path Is Also the Riskiest

Every strategic choice is a trade you make with imperfect information. Here the trade has three axes: cost, capability, and geopolitical risk. Pull hard on one and the other two shift.

Start with cost, because it is the loudest. US frontier list prices reach real money fast. Anthropic's Claude Opus 4.8 runs $5.00 per million input tokens and $25.00 per million output tokens as of June 2026. It admits a 10-point capability gap while claiming a 10x cost gap.

That admission matters. A tier of models that is 90% as good is a gift for bulk internal tasks. It is a liability for high-stakes reasoning, safety-sensitive decisions, and complex agents. The capability axis says: match the model to the job, not the hype.

Now the axis that gets waved away. Reuters reported that Chinese open-source models won wide adoption among startups but struggled to break into large businesses due to security concerns, especially in cybersecurity and finance. A model whose weights, staff, and governance sit in China is very hard for a US regulator to audit. That is a structural fact, not a slur.

Here is my read on the deeper pattern. Washington has run this play before with electric vehicles and solar panels. Tolerate the cheap import, let dependence build quietly, then move hard once the reliance becomes politically visible. If that happens, enterprises that wired core systems around Chinese cores could face forced migrations on short notice.

That is asymmetric risk. The savings are linear and show up on this quarter's invoice. The tail risk is a rewrite of your architecture under a compliance deadline you did not choose. A 20x cost cut looks brilliant right up until a procurement ban makes it a liability.

The strategic move is not "pick a country." It is counterpositioning. Route bulk, low-sensitivity, easily-portable work to the cheapest capable model, Chinese or otherwise. Keep sensitive and mission-critical work on vetted, auditable models. Build the router so switching providers is a config change, not a rebuild. Whether the current price gap survives is an open question, because US providers may reprice aggressively. Design for the world where prices move.

One more piece of the reckoning. Anthropic reportedly overtaking OpenAI in valuation and revenue signals the hierarchy itself is unstable. When the leader can change, betting your whole stack on any single vendor is the fragile choice. Optionality is the asset.

Three signals inside the same shift

2031

Pull back to the five-year arc. The question is not "who has the best model in mid-2026." It is "what does model dependence cost you when the ground keeps moving."

Goldman Sachs says global token usage could multiply 24-fold by 2030. Google said it processed over 3.2 quadrillion tokens per month in early 2026, a roughly 7x jump in a year. Volume like that means the switching decision is not one-time. It repeats every quarter as prices, capabilities, and rules shift under you.

By 2031, I expect the winning enterprises will not be the ones who found the cheapest model in 2026. They will be the ones who built the muscle to switch without pain. That is a flywheel: every provider you can route to cheaply makes you a better negotiator and a harder target for any single price hike or ban.

The contrast is stark. Amateurs optimize this month's token bill. Strategists optimize their freedom to move. Only the invoice is real today; the vendor hierarchy is just accounting, and accounting gets restated.

The data is mixed on whether Chinese models become a durable enterprise default or stay a developer arbitrage that never reaches core systems. That ambiguity is the point. In a market this unsettled, the compounding advantage goes to whoever refuses to marry a single vendor.

What to Build This Weekend

Do not rip out your stack. Build a small model router and prove the Three-Tier Reckoning on one workflow.

First, pick one high-volume task you already run. Ticket summaries. Draft emails. Log classification. Something where a small mistake costs you nothing.

Second, build a thin routing layer. A router is just an "if this, then that" traffic cop that sends each request to a chosen model. You can prototype the interface in bolt.new, which lets you build and deploy a full-stack web app in the browser, or scaffold the whole thing from one prompt in Rocket, according to its marketing. Route your test task to a cheap model and your high-stakes task to a frontier one.

Third, measure cost per business outcome, not cost per token. Run 100 real requests through the cheap model and 100 through the frontier model. Compare the bill and the error rate side by side. Boom, now you have your own "90% as good at 10% of the price" number instead of a vendor's.

Fourth, keep your provider names in one config file, not scattered across your code. Store the swap-in strings once. Use a snippet tool like typedesk to expand your standard test prompts so you run the exact same input against every model. Consistent input is the only way the comparison means anything.

Things will break. A cheap model will fumble a task you assumed was trivial. Good, that is the test working. That is data telling you which tier the task really belongs in.

You do not need a CS degree for any of this. You need one boring task, two models, and a weekend. The enterprises that win the next five years are just running this same loop at scale. Get your reps in now.