
Pulling Claude Out of Most Daily Coding, Switching to DeepSeek V4.1 Flash: AI Costs Down About 90%
Vũ Khôi
Tác giả tại Wotech Electronics Blog

Pulling Claude Out of Most Daily Coding, Switching to DeepSeek V4.1 Flash: AI Costs Down About 90%
Nguyễn Mạnh Đạt's team (behind cheapaimodel.ai) has pulled Claude out of most daily coding work. What surprised them most, according to the post in the "API không ngon xóa group" group: the work that really needs the strongest model is much less than imagined. Most of what the team asks AI every day is repetitive work — and repetitive work is work a smaller model handles well enough.
Quick answer: Before, the whole team defaulted everything to the strongest model: fixing a trivial bug meant Claude, writing a few unit tests meant Claude. After sitting down to review what the team actually asks AI each day, they moved repetitive work to DeepSeek V4.1 Flash: debugging, writing tests, everyday feature code, and running agents. DeepSeek performs fine, with good-enough quality; genuinely hard work still goes to Claude, but now that is the exception rather than the default. Technically almost nothing had to change, because the endpoint is OpenAI-compatible — just swap the base URL and model ID, keeping the old code untouched. The team's AI costs fell by about 90%, according to the author's figure.
An OpenAI-compatible endpoint is an API designed with the same calling format as OpenAI's API: the same request structure and the same way of passing parameters. Because of that, code calling one model can switch to another model just by changing the endpoint address (base URL) and the model name (model ID), without rewriting the API-calling logic.
Why did the team default everything to the strongest model?
According to the post, before, the whole team defaulted everything to the strongest model. Fixing a trivial bug meant Claude, writing a few unit tests meant Claude. It is an understandable habit: when you have the strongest model at hand, you throw everything at it to be safe, without thinking about which task needs how much power.
The turning point came when the team sat down to review what they actually ask AI each day. The review showed most of it was repetitive work: debugging, writing tests, everyday feature code, running agents — tasks with patterns, clear context, and no need for top-level reasoning ability. That finding, according to the author, was the biggest surprise: the work that really needs the strongest model is much less than imagined.
Which work moved to DeepSeek V4.1 Flash?
The team moved four groups of work to DeepSeek V4.1 Flash, and judged that DeepSeek performs fine, with good-enough quality:
| Work groups moved to DeepSeek V4.1 Flash | Role of the strongest model after the move |
|---|---|
| Debugging | Only for genuinely hard work |
| Writing tests | From default to exception |
| Everyday feature code | No longer the first choice for everything |
| Running agents | Reserved for the most demanding agent tasks |
Genuinely hard work still goes to Claude — but now that is the exception rather than the default. In other words, the team did not "drop" the strongest model; it put it in its proper place: reserving power for hard work, and letting a lighter, cheaper model handle the bulk of repetitive work that fills most of each day.
Was the switch difficult?
Almost nothing had to be done, according to the post. The endpoint is OpenAI-compatible, so just change the base URL and model ID, keeping the old code untouched. The team calls through its own API (cheapaimodel.ai), which made the switch even lighter: no logic changes, no rewritten integrations, no change in how the team works — just a different model behind the same calling interface.
This is the key point the post wants to stress: the technical barrier to trying a cheaper model is almost zero. The real barrier is the habit of defaulting to the strongest model for everything — and that habit only changes when someone sits down to measure what the team actually uses AI for.
What does a 90% cost reduction mean?
The "about 90%" figure is the author's own figure for his team, not independently verified — and it reflects that team's particular work mix: mostly repetitive work, so most of the previous spending went to the most expensive model for tasks a cheaper model could also handle. A different team with a higher share of hard work would save less.
The post closes with an open question for the community: where does each person's boundary lie — is there any task tried on a small model that had to go back to a large one? That question acknowledges something important: the boundary between "small work" and "hard work" is not fixed, and each team has to measure it on its own work, rather than following another team's experience.
Frequently asked questions
1. What did the author's team do? According to the post, the team (behind cheapaimodel.ai) pulled Claude out of most daily coding work, moved repetitive work to DeepSeek V4.1 Flash, and kept Claude only for genuinely hard work.
2. Which tasks moved to DeepSeek V4.1 Flash? Debugging, writing tests, everyday feature code, and running agents. According to the author, DeepSeek performs fine, with good-enough quality for these tasks.
3. Which tasks still need the strongest model? According to the post, genuinely hard work still goes to Claude — but now that is the exception rather than the default as before.
4. Did switching models require much code change? According to the post, almost nothing: the endpoint is OpenAI-compatible, so just change the base URL and model ID, keeping the old code untouched. The team calls through its own API.
5. Does the 90% AI cost reduction apply to every team? Not necessarily. The about-90% figure belongs to the author's team alone, reflecting its work mix — mostly repetitive work. A team with a higher share of hard work would save less, and the boundary between small and hard work has to be measured on one's own work.
Source: post by Nguyễn Mạnh Đạt in the "API không ngon xóa group" group (18/11/2026). The figures and interpretations in this article belong to the post's author and have not been independently verified by Wotech. The cover image is an illustration created by Wotech, as the original post had no attached photo.
Bài viết liên quan

18 tháng 11, 2026 · 6 phut doc
FPT Software Chairwoman: US Clients No Longer Want to Pay by the Hour, Company Shifts to Outcome-Based Pricing
Artificial intelligence is forcing FPT Software to transform its traditional software outsourcing model: large clients in the US now apply 100% outcome-based payment, and the company expects the Time…

18 tháng 11, 2026 · 6 phut doc
Meta Expands Muse to Australia: The 6th Country, Real-Time Voice, and Local App Integrations
On 17/11/2026, Meta expanded Muse to Australia — the 6th officially supported country, just over 2 months after its US launch. According to a post by Hồ Hoàng Tiến (18/11/2026), the notable point is …

18 tháng 11, 2026 · 7 phut doc
AI Agent Costs Grow With the Square of Steps: 4 Layers to Cut Token Spend Before Switching to a Cheaper Model
You don't pay for answers — you pay for how many times you re-ask the same thing. That is the central argument of Hoang Viet AI in a post in the comailo group (18/11/2026): an agent running 40 steps,…