Shiwen Dialogue Issue 25 | 'Good enough and cheap' models are rewriting whose game rules

Host: Shiwen (AI) | Guest: Yongliang

Introduction: After a month, it’s hard to tell which model is being used by its name

On October 7th, an engineer posted an article on his blog with a title that was itself a question: “Why isn’t the entire industry screaming about DeepSeek 4.1 Flash?” The next day, when it was posted on Hacker News, it received over 350 upvotes and more than 300 comments, with most discussions focusing on the scale of the savings. What really stood out was his description of how he used the model: He had heavily utilized it for a month and ran it in over a dozen projects. In conversations, without looking at the model name, he couldn’t tell whether he was using DeepSeek or Opus. Note that this is the author’s personal experience, not a benchmark test. His stance was also straightforward: I don’t care about not having 4.1 Pro; I use it as a cutting-edge model because it performs like one. To what extent it’s cheap, whether this statement is credible, and what ordinary people should do with it—these are the three main topics of this issue.

Shiwen: If a model is cheap, why is it worth discussing in a separate issue?

Yongliang: Because this time the discount isn’t “a 10% off,” it’s a different way of measuring. There’s more than 300 times the difference between $0.003 and $1. When the bill becomes this large, what changes isn’t how much you spend, but whether you dare to invest in it. “Cheap” and “good enough” need to be analyzed separately: Cheapness is about the bill, which can be verified; goodness is about the experience, which is subjective. This issue will cover both, but first we’ll talk about cheapness, treating it as a phenomenon: All facts come from the author’s own account and Hacker News comments; my judgments will be marked separately.

Shiwen: Do you believe “it’s hard to tell whether it’s DeepSeek or Opus”?

Yongliang: We’ll discuss belief after presenting the facts. Q2 will focus specifically on this, and I’ll present any skeptical comments as they are.

Q1: To what extent is it cheap? Three sets of numbers to show

Shiwen: Let’s start with the numbers. To what extent is it actually cheap?

Yongliang: Three sets of numbers, each more striking than the last.

The first set: The author used OpenCode Go subscription, $10 per month, which essentially means unlimited use by DeepSeek; his experience was that daily usage costs were rarely over $1. Note the gap between these two figures: Fixed monthly fees with almost unlimited usage, a pricing structure unimaginable a year ago.

The second set: He gave the least noticeable example, organizing desktop files. Having the model sort and name the messy files on the desk—a simple task—costs $0.003 with 4.1 Flash; the same task done by a cutting-edge model costs about $1. More than 300 times the difference, but nothing changes in the task itself; it’s only who is qualified to do it that changes.

The third set explains “why it’s so cheap”: KV cache has been reduced to about 1/437 of V1. The mechanism is simple: The longer the conversation, the greater the cost of storing in GPU memory. This reduction removes the most expensive part of long conversations. Some comments in the comment section questioned the way the numbers were written, saying “Just say it’s reduced to 1/437,” but no one questioned the numbers themselves.

Shiwen: So cheapness has engineering reasons, not just cost-cutting?

Yongliang: From the mechanism presented by the author, it seems to be an engineering effort. Cache optimization actually saves real compute power and memory. These two aspects can be compared: on one hand, some local players run it on top-of-the-line Macs with 256GB of memory, with the model fully resident in memory, pre-filling over 2,000 tokens per second and generating around 40 per second; on the other hand, some comments suggest that running it at full precision requires 1,664GB of video memory, needing eight top-tier professional cards to handle it, making self-hosting unaffordable for ordinary people. The author’s own judgment is more direct: under the economics of 4.1 Flash, “self-hosting to save money” no longer holds true. Those with real privacy concerns can wait, as cache optimization will soon be rolled out locally. The phenomenon of “the same work, at a much lower cost” has already occurred. HN has debated this for over three hundred layers, and this is why it’s worth discussing. As for how long subsidies will last, I cannot predict.

Q2: Is the statement “Can’t distinguish Opus” reliable?

Shiwen: You left a cliffhanger at the beginning. “Can’t distinguish without looking at the name,” is that enough?

Yongliang: Let’s stick to the facts first: this is the author’s personal experience. He used it intensively for a month, working on over a dozen projects, and this is his usage profile; but this statement has no backing from any benchmark tests, and he hasn’t run any benchmarks himself. Moreover, he doesn’t have no reference point—he still has frontier subscription services, and the final review of key projects still uses Opus, meaning he’s constantly comparing. Whether personal experience can be generalized to you depends on how your work resembles his. This is my reminder, not a denial. He’s honest about his own results, but his praise for the model should be labeled as “personal experience.”

The comments section isn’t one-sided either. Those who criticize are sharp: “Because it’s not that useful? It doesn’t work in serious coding scenarios,” implying its training use, but without evidence. Some are more calm: “Why should we panic? The system has always gotten better, faster, and cheaper over time”; others simply say “DeepSeek isn’t on my mind at all, frontier models are already cheap enough”; and some explain from the procurement side: “Companies move slowly, only recognizing the top two brands.” These voices coexist with the “can’t distinguish” sentiment, which is the full picture of this situation: one side complains about outrageous pricing, while the other doesn’t even consider it.

Shiwen: You’ve led a team for seventeen years. Where do you place your judgment?

Yongliang: My division is: 80% of daily tasks have “good enough” performance; the remaining 20% of critical tasks still show real differences. Writing interfaces, modifying scripts, organizing files, exploratory UI attempts—these don’t require the strongest models. What’s needed is something cheap enough to not hesitate over. For truly intense tasks (architecture decisions, security-sensitive code, handling edge cases), I still use the strongest models. This is my experience-based division, not statistics, but it matches the usage profiles in the comments: some say after a few weeks, web development is excellent across all scenarios, and the pricing is outrageous, so they cancel their original subscription; others cancel their more expensive Max tier subscription, saying in certain scenarios it’s even better and faster than Opus 5.5, and they expect to be able to run it on laptops within a year; some keep the expensive tier and add that “they won’t need it after another iteration.” Keeping the option open shows that the 20% is still effective.

Q3: Dumbbell strategy—cheap for daily tasks, expensive for final reviews

Shiwen: Is the correct approach “switching all to the cheaper ones”?

Yongliang: It’s not a complete replacement, it’s like a dumbbell: thick at both ends and thin in the middle.

The author’s workflow is worth copying: all daily tasks are handed over to 4.1 Flash, while boring tasks and exploratory attempts are sent out without any psychological burden. At critical points, ask Opus 5.5 for final review, focusing on edge cases, and then have DeepSeek execute the fixes. His conclusion is more valuable than the process itself: for key tasks, “a new perspective” is more useful than “a stronger model.” The value of final review lies in providing a different perspective, not increasing power.

There’s a team on HN that follows an enhanced version of the same logic: all personal assistants run on Flash by default, while hard tasks are handled by professional sub-agents, with heavier work assigned to stronger models. Cheap models act as schedulers, while expensive ones act as experts—this is also a dumbbell.

Shiwen: What inspired you when selecting teams you’ve led? Yongliang: One word: frontier tax. Looking back at the years I’ve led teams, how much of the budget was spent on “getting the best”? Eighty percent of tasks were just good enough. High-end models are right, but the mistake was letting them do everything, like equipping every seat with a sports car for delivery. Why cut from the middle tier? Because both ends can clearly answer “why pay”? The top-tier answers quantity, the final review tier answers risk; only the middle tier doesn’t rely on either: being less reliable than the strongest tier, but still more expensive than the good-tier ones. The essence of a dumbbell is paying based on task difficulty: cheap for the main tier, expensive for the final review tier, and the “middle” tier’s overhead should be cut. This is my experience, not a law.

Q4: $0.003 in psychology costs less than the threshold, and behavior changes.

Shiwen: You started by saying “It’s not how much money is spent, but whether you dare to assign tasks.” Tell us more. Yongliang: The meaning of $0.003 isn’t in saving the remaining $0.997, but in falling below a psychological threshold.

In the era of high-end models, before assigning tasks, there was always this thought: “Is it worth spending one dollar asking AI?” Three times out of ten, you’d calculate and endure. After lowering it to $0.003, this thought disappears. It’s not that you become more decisive, but that the decision cost itself disappears. The author has a harsh statement: “Asking a mathematician to organize desk files is an insulting use of models.” Reading it backwards means a new normal: cheap enough to insult them, and usage is liberated. Someone in the comments used AI to make 25 small items, spending a total of $0.70. It’s not that they’ve become more diligent, but that “trying one” no longer has a cost. By the way, the author also has a personal inference: cache optimization means less power and water, he says that now when opening Claude, “it feels almost wasteful.”

This scenario has happened before. In the early days of cloud computing, when pay-as-you-go just appeared, many people calculated that “renting by the hour was more expensive than buying servers themselves.” Later, everyone realized it wasn’t about saving money: no need for advance approval for purchases, no need to pay for peak usage, errors could be easily discarded, and the entire infrastructure habits changed. Pay-as-you-go changed not just the bill, but the organization’s courage. Models being priced at $0.003 per task is the second performance of the same scenario.

Shiwen: So who exactly “rewrites the rules of the game” in the title?

Yongliang: Two groups of people who live within the “highest budget”. One group is users: those who still use cutting-edge model pricing for tool budgets, paying 100% for 20% of their needs. The other group is those that sell at a premium based on “best”. When 80% of work is taken by cheaper models, the pricing power of “best” is only maintained for those two percent of scenarios. The same applies to individuals: paying monthly subscriptions is essentially paying at “best”; in a world where work is priced per piece, the same task is clearly marked with a price. When these two accounts are put together, the extra amount is actually buying peace of mind. This is my analysis, not statistics.

Conclusion

Shiwen: What’s your summary for this issue?

Yongliang: $0.003 is not a discount, but a new unit of measurement. When a model becomes so cheap that it can be used freely, “good enough” turns from compromise to configuration: the main part is cheap, the final review is expensive, saving the middle ground. The author’s original words are suitable for this conclusion: chasing the latest and best is foolish. If you really want to pursue something, focus on one problem: for the same task, am I paying ten times more?

Shiwen: Billing doesn’t lie, especially not when it lies kindly for you. See you next time.

【Deep Dive】Free Advice: Audit Your AI Billings Tonight

After discussing phenomena in this issue, I’ve given readers three free steps they can complete tonight.

Step one: Collect bills. Bring out all your AI subscription and usage bills from the past month, dividing them into two categories: “daily work” (writing, editing, organizing, querying, testing) and “critical reviews” (important decisions, sensitive code, final review). Most people’s bills will look ugly: 80% of money is spent on critical reviews, while work is done hesitantly, which is the opposite of the dumbbell strategy.

Step two: Calculate unit price. Pick ten daily tasks you did last week and ask yourself: If this task was priced at $0.003 each, how much should have been spent this month? Compare this with the actual subscription fee. The difference isn’t about immediately canceling your subscription; it’s about understanding how much “cutting-edge tax” you’re paying. Then ask yourself: If tomorrow the main model price increases tenfold, will your workflow collapse? Yes, that means you’ve made cheapness a prerequisite. Cheapness can be enjoyed, but don’t bet everything on it. The saying “chasing the latest and best is foolish” should be posted next to your monitor.

Step three: Balance the dumbbell. If the main model becomes cheaper, don’t upgrade unless it’s sufficient; keep a more expensive model for critical reviews, setting boundaries and getting second opinions. Whether to cancel your subscription isn’t important. The person in the comment section who keeps Claude’s subscription and says “I won’t need it after another iteration” is the healthiest stance: keep a way out, but know why you’re keeping it.

Finally, one cold water note: $0.003 is today’s price, not eternal; “not being able to distinguish Opus” is the author’s personal experience, not something you can also do with your tasks. Cheapness changes behavior, but don’t let cheapness replace judgment. You can’t play without either end of the dumbbell.

Sources for This Issue

• Blog “Why isn’t the industry freaking out about DeepSeek 4.1 Flash?” (Jono’s Corner, dgt.is, 2026-10-07, original article with firsthand views): One-month intensive use and over a dozen projects, $10/month OpenCode Go subscription, very few daily conversations exceeding $1, $0.003 vs. roughly $1, KV cache reduced to about 1/437 of the previous generation; Opus 5.5 final-review workflow, “New perspectives are more useful,” “ chasing the latest and strongest is foolish,” “insulting usage” all出自 this article.
• Hacker News story 50000488 (378▲/305💬, reading on 10-09, over 350▲ level of intensity): The cold-water faction said “serious coding scenarios don’t work,” “why panic,” “not considered,” feedback on canceling subscription, 25 cases in a day totaling $0.70, team’s default Flash with professional sub-agent added, “written as one-fifth of 437” complaints, all relayed in the comment section, no names mentioned.
• Guest’s personal experience (17 years of technical management, team selection “frontier tax,” 80/20 experience segmentation, cloud computing pay-as-you-go analogy): First-person narrative, not statistical.
• Tone and redline implementation: “Can’t distinguish Opus” all references include “author’s own statement” tone; geopolitical comments, qualitative attribution of training data, stock price and vendor landscape topics not included in the main text.