Making $3,000/Month Freelancing with Codex? A Rational Look at AI Coding Side Hustles

A "Wealth Story" That Keeps Getting Repackaged

Recently, a wave of irresistible headlines has been flooding platforms like Bilibili: "I made ¥24,000 my first month doing coding gigs with Codex after work" and "28-year-old with zero coding background struck gold in just one month with AI programming." The narrative is remarkably consistent — zero barrier to entry, high returns, all you need is a laptop — complete with "full learning packs" and "freelance platform guides."

As tech practitioners, we owe it to ourselves to critically examine these claims: which parts reflect real trends, and which are pure marketing? AI coding tools genuinely are lowering the barrier to software development — but the claim of "zero-to-¥20K/month in one month" deserves serious scrutiny.

The Four-Week Learning Roadmap: Useful Framework, Unrealistic Expectations

Let's start with the four-week learning roadmap itself. Strip away the marketing, and the underlying framework isn't entirely unreasonable:

Breaking Down the Four-Week Plan

• Week 1: Get familiar with Codex's four core capabilities — requirement decomposition, code generation, project comprehension, and error fixing. This is the foundation phase: understanding what an AI coding assistant can and cannot do.
• Week 2: Master the standard AI-assisted development workflow and "prompt thinking" — how to get the model to autonomously break down complex tasks, covering writing, modifying, and debugging code end-to-end.
• Week 3: Level up to AI Agent development, including tool calling, API integration, task planning, and automated execution.
• Week 4: Ship real-world projects, such as intelligent customer service bots, personal knowledge bases, office automation tools, and AI-powered data analysis.

Technical background on Codex: OpenAI Codex is an AI system built on the GPT family of large language models and specifically fine-tuned for coding tasks. Its training data includes tens of billions of lines of public code from GitHub, and it served as the underlying engine for GitHub Copilot. Architecturally, Codex is a specialized variant of an autoregressive language model — through supervised fine-tuning (SFT) on massive code corpora, the model develops strong biases toward code syntax, API call patterns, and programming logic during token prediction. This technical approach established the "general language model + code fine-tuning" paradigm that was later adopted by open-source code models like Code Llama and DeepSeek Coder. Codex was originally released in 2021, based on the GPT-3 architecture and fine-tuned specifically for coding tasks across dozens of programming languages on GitHub. It's worth noting that OpenAI shut down the standalone Codex API in March 2023 — its code generation capabilities have since been integrated into GPT-4 and the subsequent GPT-4o series. With the introduction of longer context windows (up to 128K tokens), code generation evolved from "snippet completion" to "cross-file project comprehension," enabling AI to simultaneously understand dependencies between frontend components, backend APIs, and database schemas. GitHub Copilot also migrated to this more powerful underlying engine. Since 2023, with the widespread adoption of more capable models like GPT-4 and Claude 3.5, the concept of "AI coding assistant" has expanded well beyond code completion to encompass requirements analysis, architectural recommendations, and test generation — further reducing the technical barrier to software development.

Objectively speaking, the technical logic of this learning path is sound: moving from tool familiarity, to workflow, to Agent capabilities, and finally to real projects. Anyone looking to systematically learn AI-assisted development can use this kind of progressive structure as a reference.

The AI Agent topic covered in Week 3 represents one of the fastest-evolving AI application paradigms in recent years. Its core architecture typically consists of four modules: a perception layer (receiving user instructions or environmental signals), a planning layer (where an LLM generates task decomposition plans), a tool layer (calling external APIs, code executors, databases, etc.), and a memory layer (short-term context and long-term knowledge storage). The central idea behind Agents is to move large language models beyond passively answering questions — enabling them to actively plan tasks, invoke external tools, and iteratively work toward completing complex goals. Well-known frameworks include LangChain, AutoGPT, and Microsoft's AutoGen.

In engineering practice, however, Agents face challenges that far exceed what demo scenarios suggest. From a systems design perspective, a production-grade Agent must solve several core engineering problems: idempotency design (ensuring that retrying a task after a network interruption doesn't produce unintended side effects), state persistence (saving task progress across sessions to prevent context loss from breaking the task chain), and error isolation in tool calls (a single tool failure shouldn't crash the entire task). Hallucination-induced planning errors — where the model generates execution steps that seem plausible but are actually incorrect — compound exponentially in multi-step automation scenarios. The industry metric "Task Success Rate" reveals a sobering reality: even GPT-4-level Agents working on complex tasks involving more than five tool calls often achieve end-to-end success rates below 60%. Additionally, managing permissions and security boundaries for tool calls involves backend engineering practices such as API key security, rate limiting, and circuit breakers. Cost control in production environments (GPT-4-level API call costs can accumulate rapidly in complex Agent tasks, requiring careful token budget management in commercial projects) all demand solid backend engineering and prompt engineering experience to handle effectively. This is precisely why the claim of "leveling up to Agent development in one week" is technically over-optimistic.

The problem is that compressing this entire roadmap into "commercially deliverable in one month with zero prior experience" severely underestimates the complexity of software engineering.

Where Does the "¥24,000 First Month" Claim Fall Apart?

Delivery Capability ≠ Being Able to Generate Code

One line in these videos actually gives the game away: "The core of AI coding isn't knowing how to prompt — it's being able to deliver." That statement is correct — and it's precisely where the difficulty lies.

AI can help you generate a chunk of code, but delivering a commercial project involves requirements gathering, architectural design, error handling, deployment, operations, and ongoing maintenance. These are exactly the things someone with zero experience cannot master in a month. There is a massive gap between getting Codex to generate a demo and delivering an application that runs reliably and that clients will pay for on an ongoing basis.

The Real Prices and Competition in the Freelance Market

These videos claim "small tools start at a few hundred yuan, enterprise apps go into the tens of thousands." But the actual freelance market tells a very different story.

The global freelance developer market is shaped by two forces: information asymmetry and platform mechanism design. From an economics standpoint, platforms like Upwork and Fiverr are fundamentally two-sided markets: they must simultaneously maintain engagement from both project posters (demand side) and freelancers (supply side), with rating systems serving as the core mechanism for resolving information asymmetry between buyers and sellers. The systemic challenge facing newcomers is known as the "cold start problem" — no ratings means difficulty getting jobs; no jobs means no ratings to accumulate. This cycle means new user order conversion rates on these platforms typically fall below 5%. Upwork data shows that developers charging more than $50/hour typically have 3+ years of verifiable delivery history; premium platforms like Toptal claim to accept only the top 3% of applicants. Developers from India, Pakistan, the Philippines, and similar regions hold a clear price advantage on Fiverr due to lower cost of living. Domestic Chinese platforms like Zhubajie, Chengxuyuan Kuzhan, and Mashi show similar dynamics: competition for sub-¥1,000 requests is fierce, while projects over ¥10,000 typically require the freelancer to provide a business license, sign a formal contract, and agree to penalty clauses.

At a deeper structural level, the widespread adoption of AI tools has created a "skill commoditization" effect: when large numbers of participants can complete similar tasks using similar tools, price competition becomes the primary differentiator, driving overall platform price levels downward. The true value premium shifts to higher-order capabilities that AI tools cannot replace — including cross-domain business understanding, long-term client relationship management, and holistic architectural design for complex systems.

• Sub-¥1,000 small tool requests face intense competition with consistently suppressed rates — hardly enough to support a ¥20,000/month goal;
• For enterprise-level projects worth tens of thousands, clients typically require proof of qualifications, case studies, and post-delivery support — nearly impossible for a complete beginner to provide.

The "¥24,000 first month" figure is far more likely to come from selling courses and resource packs, not actual coding freelance income. This is the most important red flag in this type of content. Content creators attract traffic by showcasing high-income scenarios, then convert that traffic into course sales, resource bundles, or community membership fees. This commercial loop is known in the knowledge economy as the "selling shovels" model — a reference to the California Gold Rush of the 19th century, when the people who reliably made money weren't the miners rushing for gold, but the merchants selling shovels, tents, and food to those miners. What makes this model especially notable is that its profitability is naturally decoupled from the actual income potential of the skills being taught — more course sales means more revenue, regardless of whether students can actually make money through coding gigs. Platform algorithms' tendency to amplify emotionally engaging content further reinforces the spread of "fast-track income" narratives. China's knowledge-payment market exceeded ¥60 billion in 2023, yet according to third-party surveys, fewer than 10% of buyers successfully monetize what they've learned. An effective way to evaluate this type of content is to ask whether the income source is verifiable — genuine freelance income can be demonstrated through platform transaction records and contract screenshots, while course sales revenue has nothing to do with the actual income potential of the skills being taught.

Where Are the Real Opportunities in AI Coding?

Strip away the marketing shell, and AI-assisted coding genuinely is a direction worth taking seriously. Rising demand for AI talent is an objective trend — companies of all sizes are increasingly looking for engineers who can skillfully use AI tools to boost R&D efficiency.

Recommendations by Audience

Developers who already have a coding foundation: Tools like Codex, Claude, and GitHub Copilot can significantly boost your development efficiency and are well worth deeply integrating into your daily workflow. GitHub's own data shows that developers using Copilot can see efficiency gains of up to 55% on specific coding tasks — a genuinely meaningful competitive advantage.

Beginners who want to break into tech: AI has lowered the barrier to entry, but it cannot replace systematic learning. Treat AI as "a smarter textbook and practice partner" — build a solid understanding of programming fundamentals, data structures, and deployment workflows, rather than fantasizing about skipping all of that to start making money right away. The core concepts of computer science — variables and data types, control flow, functional abstraction, object-oriented design — are cognitive foundations that AI tools cannot help you bypass. Understanding them deeply is what allows you to evaluate whether AI-generated code is correct and reliable, and to intervene effectively when the output goes wrong.

People who want a coding side hustle: First build a real project portfolio through genuine experience, start with smaller requests, and gradually establish trust through delivery quality and reputation — not through a "fast-track bootcamp." A developer with real commit history on GitHub and a complete review history on a freelance platform is far more persuasive than any course certificate.

Thinking Clearly About "Fast-Track Income" Promises

This type of content spreads widely because it precisely targets the anxiety of "low barrier + high return + after-work side income." The "availability heuristic" — systematically described by Nobel laureate Daniel Kahneman in Thinking, Fast and Slow — reveals that the human brain tends to assess probability based on how easily information comes to mind rather than statistical reality. The specific number "¥24,000 first month" and emotionally charged narrative structures are classic triggers for this cognitive bias.

Platform algorithms also play a deep supporting role in this dynamic. Beyond the availability heuristic, the "social proof" principle is equally at work: when a video showcases large numbers of likes, shares, and "verified it works" comments, audiences interpret group behavior as evidence of information reliability — even if those interactions themselves may have been artificially amplified through operational tactics. From a content production business logic standpoint, specific figures like "¥24,000 first month" significantly increase the likelihood that viewers watch the entire video, because the brain encodes specific numbers far more efficiently than abstract descriptions — boosting completion rate and engagement rate, the two metrics that matter most to recommendation algorithms. The psychology of loss aversion also plays a role here: when content implies that "not jumping in now means missing the AI wave," audiences' fear of potential loss tends to override rational evaluation of potential gains. This is why people systematically overestimate the prevalence of vivid, specific success stories while ignoring the silent majority of failures.

But there are no shortcuts when it comes to building technical skills. AI tools are fundamentally amplifiers — they amplify the efficiency of those who already have a foundation, but they cannot conjure delivery capability out of thin air.

If you're genuinely interested in AI coding, the right approach is: treat this type of content as a starting point for understanding trends, not as a reason to pay for a course. OpenAI, Anthropic, and Microsoft's official documentation, free learning communities like freeCodeCamp and The Odin Project, and the tens of thousands of open-source projects on GitHub are more than enough to support systematic learning. What's truly scarce was never access to materials — it's the sustained time investment and the patience to actually build things.

For any content promising "zero background to ¥20,000/month in one month," maintaining a healthy dose of skepticism is the basic critical thinking every practitioner in tech should exercise.

Key Takeaways