Build Anything with OpenAI o3, Here’s How

OpenAI’s o3 (and o4-mini) is presented as a step-change in what AI can do end-to-end: it can reason at “PhD level,” zoom into and analyze images or PDFs by selectively inspecting regions, and then use tools—up to hundreds of tool calls—to browse, code, and run tasks autonomously in a terminal. The practical takeaway is that these models aren’t just good at answering questions; they’re being positioned as builders that can take a vague goal (like finding and completing paid Upwork work) and turn it into working software with minimal human intervention—provided the workflow is set up correctly.

The transcript first grounds the claim with benchmarks. On AME 2024 (competition-level math), o4-mini reportedly outperforms o3. Coding performance shows large jumps across generations, with a cited score of 2700 on Codeforces “like top 200 programmers.” For real-world project completion, the SWE-Lancer benchmark is used to argue that o3 and o4-mini can complete Upwork-style projects worth significantly more than earlier models. Multimodal capability also improves, and o3 is described as stronger at editing code—generating entire files, updating specific sections, following instructions, and operating in an “agentic” way.

A key mechanism behind the gains is attributed to reinforcement learning scaling: OpenAI allegedly found that more compute leads to better performance, applying the same lesson from earlier GPT-style models to reasoning-focused “o” models. The transcript claims that at the same latency and cost as o1, o3 delivers higher performance, and that adding more compute continues to raise results—pushing back against claims that AI progress is slowing.

The middle portion demonstrates multimodal power with image analysis. o3 is shown identifying a specific location on Earth from a screenshot, using zoomed-in inspection of different image regions during reasoning. Another example involves a physics paper where the model zooms into relevant parts—charts, paragraphs, and sections—then synthesizes conclusions. The implication is that contract-style documents could be uploaded and analyzed section-by-section, with relationships between clauses used to produce detailed answers.

The latter half turns to a money-making workflow: using o3 to browse for suitable Upwork listings and then using Codex CLI (an open-source autonomous coding agent that runs in the terminal) to implement the work. The process starts by feeding o3 the Codex CLI repository README and release information, leveraging o3’s large context window (200,000 tokens) so it can plan with tool capabilities in mind. The transcript emphasizes that o3 and o4-mini can access “all tools” available in the environment—web search, file analysis, Python execution, visual reasoning, and image generation—before producing an answer.

A concrete example targets a Git-related Upwork project (“git commit to change log CLI”). The transcript then walks through setting up Codex CLI on a MacBook: installing dependencies (including Homebrew and Node checks), installing the Codex CLI package via npm, creating an OpenAI API key, and running smoke tests. It highlights a practical reality: autonomous coding can fail when the working directory isn’t a Git repo, when API keys or environment variables aren’t wired correctly, or when build tooling expects specific versions. Errors are handled by iterating—asking o3 to diagnose and fix issues, re-running commands, and sometimes re-initializing git.

Finally, the transcript shows building a “fanciest to-do list app” in full auto mode, then attempts a more complex Upwork-style GitHub project that generates commit reports. When Codex CLI struggles, the workflow shifts to using Cursor as an IDE and o3 as the planning/advising layer, with step-by-step “project scope” instructions executed incrementally. The end result is a working MVP that summarizes commit activity into markdown reports, plus a “nice to have” checklist (README and permissive license). The broader message is that with the right toolchain, o3 can compress the cycle from idea → implementation → runnable software, making paid project delivery more feasible for individuals—though still requiring careful setup and iterative debugging.