- Authors
- Name
- 오늘의 바이브
Praise Goes to Opus, Work Goes to Codex
In February 2026, Fortune ran an interesting piece. OpenAI had launched a desktop Codex app. But buried in the article was a notable figure: Peter Steinberger, creator of the viral AI tool OpenClaw.
Steinberger publicly called Anthropic's Claude Opus "the best for general-purpose agents." The best. Full stop. And yet he built all of OpenClaw using OpenAI's Codex. He even said his productivity roughly doubled.
Why? If Opus is the best, just use Opus. He praises Opus but does the actual work with Codex. Is this a contradiction, or is something else going on?
What Is OpenClaw
First, some context. OpenClaw is an AI-powered development tool built by indie developer Peter Steinberger. It automates code generation, refactoring, and debugging. It went viral almost immediately, spreading through developer communities by word of mouth.
The core of OpenClaw is an AI agent-based workflow. You give it a task, and multiple AI agents collaborate to handle it. One agent analyzes the code, another suggests fixes, another runs tests. Each has its own role.
Steinberger built this alone. A solo developer created a complex AI agent system. How? He said it himself. He used Codex.
"Best" and "Most Productive" Are Not the Same
Look at Steinberger's statement again. "Claude Opus is the best for general-purpose agents." The keyword here is "general-purpose."
What does general-purpose mean? Broad reasoning, complex logic, creative writing, nuanced analysis. Opus excels at these tasks. In benchmarks, Opus consistently tops the charts for complex reasoning, math, and coding challenges.
But real-world development work might not be "general-purpose." Writing code is closer to a specialized task than broad reasoning. Pattern generation, API reference lookups, error fixing, refactoring. For these, "the model optimized for the job" can outperform "the smartest model overall."
Codex, as the name suggests, is built for coding. OpenAI designed it for developers. It is optimized for code generation, completion, and explanation. It may trail Opus in general reasoning, but in the specific domain of writing code, it can be faster and more efficient.
Steinberger's claim of "doubled productivity" makes sense in this context. He did not pick the smartest model. He picked the most efficient one. Praise to Opus, work to Codex. Not a contradiction. A rational division of labor.
How to Choose a Tool — Intelligence vs Efficiency
When developers pick an AI tool, they weigh several factors.
1. Accuracy Does it generate correct code? Does the output actually run without errors?
2. Speed How fast are the responses? How much latency builds up during repeated use?
3. Cost How much per API call? Is it economical at scale?
4. Integration How well does it fit existing workflows? Is it easy to connect with your IDE, terminal, and CI/CD?
5. Consistency Does it produce consistent results for the same request? Is the output predictable?
Even if Opus leads in accuracy and raw intelligence, Codex can win on criteria 2 through 5. For repetitive development work, speed, cost, and consistency are decisive.
Building OpenClaw must have been a cycle of repetitive tasks. Adding features, fixing bugs, refactoring, testing. Each time, you prompt the AI and review the output. In that loop, even a small speed advantage dramatically improves perceived productivity.
| Criteria | Claude Opus | Codex |
|---|---|---|
| Accuracy | Very high | High |
| Speed | Moderate | Fast |
| Cost | High | Relatively cheaper |
| Integration | Moderate | High (IDE) |
| Consistency | High | Very high |
Steinberger understood this tradeoff. He chose peak efficiency over peak intelligence.
The Reality of AI Tool Competition
According to the Fortune article, OpenAI's Codex has over one million monthly active developers. The desktop app launch will push that number higher. The new app lets you run multiple AI agents simultaneously, automate repetitive tasks, and monitor agent activity.
Meanwhile, Anthropic's Claude Code is growing fast. Adoption is rising at major companies like Uber, Netflix, and Spotify. Claude Code is terminal-based, understands entire projects, and performs complex tasks autonomously.
The two tools take different approaches.
Codex App:
- Desktop-based (separate from traditional IDEs)
- Runs multiple agents simultaneously
- Emphasizes repetitive task automation
- Recommends the GPT-5.2-Codex model
Claude Code:
- Terminal-based (developer-friendly)
- Understands full project context
- Performs tasks autonomously
- Supports Opus/Sonnet model selection
It is hard to declare one "better" than the other. It depends on context. Steinberger's choice proves this. He praised Opus as the best, but built with Codex.
The Trap of "Best"
The word "best" should be used carefully. Drop the context and it breeds misunderstanding.
"Opus is the best" probably means "under certain conditions, for certain tasks, measured by certain criteria, it is the best." No tool is the best at everything for every situation.
Developers know this. That is why they mix tools. Ask Opus about complex architecture decisions. Run Codex for repetitive code generation. Summarize papers with Claude. Analyze data with GPT. Each tool has its strengths.
Steinberger likely did the same. He acknowledged Opus as "the best" while recognizing that Codex was a better fit for his specific work. That is not a contradiction. That is expert judgment.
Breaking Free from Tool Lock-in
The AI tool market moves fast. Today's "best" has no guarantee of being the best tomorrow. In 2023, GPT-4 was dominant. In 2024, Claude 3 arrived and shifted the landscape. In 2025, Gemini climbed the ranks. In 2026, the competition is fiercer than ever.
In this environment, locking into a single tool is risky. The mindset of "this is the best, so I will only use this" kills flexibility.
Smart developers learn multiple tools. They map out each tool's strengths and weaknesses. They pick the right one for the task. When a new tool drops, they evaluate it quickly and decide whether to adopt it.
Steinberger praising Opus while building OpenClaw with Codex reflects this flexible thinking. He is not loyal to one tool. He picks the one that fits the job.
Why Publicly Praise Opus at All
Here is a fair question. If he built OpenClaw with Codex, why go out of his way to publicly praise Opus?
A few possibilities.
1. He meant it He genuinely believes Opus has superior general reasoning. Codex just happened to fit his specific workflow better.
2. Balanced evaluation In the AI tool community, picking sides makes you look biased. Acknowledging multiple tools' strengths builds credibility.
3. Playing the long game As OpenClaw grows, it will need to support multiple AI models. No reason to create adversarial relationships with any company.
4. Technical respect Developer culture respects acknowledging good technology, even from competitors. He recognized Opus's technical excellence while choosing Codex for its practicality.
Whatever the reason, it shows that "calling something the best" and "actually using it" are separate things. We often confuse the two. We assume that if someone says a tool is the best, they must be using it. Reality is different.
Pragmatism in AI Tool Selection
The takeaway from this story is pragmatism. When choosing AI tools, finding "the best" matters less than finding "the best for your situation."
What does your situation look like?
Nature of the work Do you need complex reasoning, or repetitive generation? Creative output, or consistent output?
Frequency of use How many times a day do you use it? How many requests per session? High frequency makes speed and cost critical.
Integration environment What IDE do you use? What is your workflow? A tool that fits your existing environment boosts productivity.
Budget How much can you spend on API calls? What is the team size? You need to calculate cost-effectiveness.
Team capabilities What tools is your team familiar with? Can you invest time learning something new?
Factor all of this in, and you will see not "the best AI" but "the right AI for us." For Steinberger, that was Codex.
Competition Benefits Everyone
The competition between Opus, Codex, Claude Code, and other AI tools benefits users. Competition drives each company to deliver better features, faster speeds, and lower prices.
In early 2025, AI coding tools were assistants. Now they are becoming core infrastructure. According to the Fortune article, some developers say they have "given up traditional programming." AI writes most of the code. Humans just set the direction and review.
This shift was powered by competition. A monopoly would have slowed progress. OpenAI, Anthropic, Google, and Meta competing fiercely is what drives rapid advancement.
For users, having more choices is a blessing. If Opus fits, use Opus. If Codex fits, use Codex. Switch between them as the situation calls for it. Flexible, not locked in.
Steinberger's Real Message
Let us revisit Steinberger's statement.
"Claude Opus is the best for general-purpose agents. But I built OpenClaw with Codex, and my productivity doubled."
What is the real message?
Tool selection depends on context. Pick the optimal tool, not the best tool. General-purpose performance and specialized performance are different. Praise and selection are separate decisions.
Know multiple tools. Stubbornly sticking with one tool narrows your view. Someone who understands both Opus's strengths and Codex's strengths makes better choices.
Pragmatism wins. Do not choose because "it is the best." Choose because "it fits my situation."
What Is Your "Best"?
If you are a developer using AI tools, ask yourself.
What do you think is the "best" tool? Is it actually optimal for your work? Or are you using it just because everyone says it is the best?
Have you tried other tools? Have you compared them? Or have you settled on one and stopped looking?
What is the nature of your work? More complex reasoning, or more repetitive generation? Is speed more important, or accuracy? Are you cost-sensitive, or feature-sensitive?
Working through these questions will reveal "the right tool for your situation." It might be Opus. It might be Codex. It might be Claude Code. Or it might be something else entirely.
The point is to stop chasing the "best" label. Try things yourself, compare, and evaluate in your own context. Like Steinberger did.
Not a Contradiction. Wisdom.
Back to the original question. "If Opus is the best, why did he build it with Codex?"
Now you know the answer. It was not a contradiction. It was a wise choice.
The best general intelligence and the best practical efficiency are different things. Acknowledging both while picking the one that fits your work is smart. Praise to Opus, work to Codex. That is not wrong. That is right.
In the age of AI tools, what we need is not an obsession with finding "the best." It is the discernment to pick "the optimal." Knowing multiple tools, understanding each one's strengths, and choosing flexibly based on the situation.
Steinberger showed us that. OpenClaw's success is the proof. He did not pick the tool everyone praised as the best. He picked the tool that doubled his productivity. And the results proved him right.
What about you? Are you still chasing "the best," or are you already using "the optimal"?
Sources: