Claude Code vs Codex: Which AI Coding Agent Is Better for Real Work?

Neither wins outright. After two weeks running Fable 5.1 in Claude Code and GPT 6 Astra in Codex on the same product, we found they are good at different halves of the job. Fable is the stronger builder: it produces working UI from a vague prompt. Astra is the stronger reviewer: it reads the code instead of the author's story and finds the things the builder could not see. The combination beats either one alone, and the reason is not that one model is smarter. They are trained on different data, so they are wrong about different things.

This is not a benchmark. It is what happened while we built SentiSense, a market data platform for investors and AI agents, with both agents working the same tickets every day.

What is the difference between Claude Code and Codex?

Both are terminal agents: you describe the work, they read the repository, write code, run tests, and hand back a change. Claude Code is Anthropic's, Codex is OpenAI's. On paper they do the same thing. With Fable and Astra, the difference shows up in temperament. Fable fills in what you did not say and usually guesses well; ask for "a calmer pricing button" and you get a design, not a question. Astra wants the spec and does exactly that, which makes it worse at the blank page and better at the audit.

Which one is better at building UI?

Fable, and it is not close in our experience. The header offer on our app, the pricing page, the earnings board, the story pages: each started as a one-line request and came back as a finished screen with sensible spacing, dark mode, and tests. Astra produced correct components but needed the layout described in detail, and it tended to ship a generic look until told exactly what to change. When we did hand it a CSS polish pass with precise instructions, the result was clean and accessible. It executes; it does not improvise.

UI is where taste lives, and taste is subjective. That is a reason to pick one agent for it and let that agent accumulate your preferences, rather than splitting the work and getting two half-consistent design languages.

Why does the AGENTS.md vs CLAUDE.md split force you to pick a builder?

Because the two agents do not read the same instruction files. Codex reads AGENTS.md. Claude Code reads CLAUDE.md and its own memory folder, and as of this writing it does not read AGENTS.md unless you import it. Shopify's CEO Tobi Lütke said in August he was considering banning Claude Code over exactly this: on a large monorepo some directories end up with one file and not the other, different engineers use different tools, and the ones on the wrong tool work, in his word, with a lobotomy.

For a small team the practical effect is that the split decides your roles. Whichever agent's files hold your rules, skills and memory becomes the builder, and it should be the one whose taste you prefer, because a builder that has read a year of your design notes will make the call you would have made. Ours is Claude Code, so the repository carries CLAUDE.md files and Claude's memory. The second agent starts cold every time. That is acceptable for a reviewer, and it is part of why the reviewer catches things: it reads the plan and the diff you hand it, not the accumulated story. But it means the handoff has to carry the context explicitly. We keep a root AGENTS.md whose first instruction to Codex is to read the nearest CLAUDE.md before touching a subsystem. Without that one file, the second opinion really is lobotomized.

Can Claude Code build an iOS app?

Yes, and this is the case that convinced us. In late June we parked a React Native prototype and started a native Swift app from scratch. Claude Code wrote it: 96 commits, about 14,000 lines of Swift, five Siri intents so you can ask your phone for a stock's sentiment or the market mood without opening the app, a three-tab interface, home screen and Control Center widgets, and StoreKit billing. The first TestFlight build went up within a week. Thirty-eight TestFlight builds later it is a few weeks from the App Store. It is not a demo; it is a product line that did not exist in May, and it is coming soon.

The part worth stressing for anyone skeptical of AI-written code: every build ran through the same gates as our human-written Java. Compile, unit tests, a review pass from the other agent, and a person deciding whether to ship.

Can an AI agent do stock analysis?

It can do the work of analysis, which is different from having the answer. We run curated research reports on about 970 stocks, and the agents write and refresh them from filings and verified fundamentals. What makes that trustworthy is a second pass: a fact-check skill that grades every financial claim as verified, discrepancy, stale, unverifiable, or fabricated, and a rule that a report with an unverifiable claim does not ship. The same research workflows ship as agent skills that a Claude Code, Codex or OpenClaw agent installs in one line; the publisher page on ClawHub lists 21 of them at about 18,900 installs as of this writing, which is a useful check that other people's agents run these workflows too, not only ours.

The other half is that we use our own tools. The MCP connector and the skills we publish are the same ones our agents use for our research: a Claude Code or Codex session pulls sentiment, analyst targets, insider and institutional flows, options positioning and filings through the same endpoints a customer's agent would, and that is how the reports on those 970 stocks get refreshed. The practical change is coverage and discipline. An agent reads every filing and every analyst note before it writes a sentence, so the research covers more names than a small team could by hand, and the fact-check pass refuses any claim it cannot source, which is the kind of unforced error a rushed human analyst lets through.

Which one is better at code review?

Astra, clearly, and this is where the second agent pays for itself. A few examples from one week, each on code that was about to ship:

• Claude Code wrote a fix for after-hours prices and claimed it added one extra call per ticker per night. Codex traced the actual serving paths and showed the bound was false, then found a second bug where a morning earnings reaction was filed under the wrong session.
• A comparison between two product-matching parsers reported how often they agreed. Codex found that build failures were being logged as agreement, so the parity number could not go down. Claude Code's fix reverted to the old code under new tests and showed all six fail.
• A skill that reports company growth metrics could return plus 50 percent or plus 25 percent for the same quarter depending on which duplicate row arrived first. Caught before release.
• Our post-deploy check fetched the pricing page with curl and grepped for the offer. Codex pointed out the offer renders on the client, so the check was blind and would have reported a failed deploy that had succeeded.

Astra is also unusually gracious about it: its reviews credit what the other agent got right before listing what it got wrong, which makes the verdicts easier to act on.

Should you use both?

If the work matters and the budget allows a second agent, yes, with a role for each. The case for two is not that one is smarter. They are different models trained on different data, so they see the same codebase slightly differently, and a review is the one place where a different blind spot is the whole point. Let the builder build and the other model review the diff, not the summary. Ask the reviewer for a verdict with a shape: ship, fix then ship, do not ship. And make the builder prove the fix with a failing test rather than a paragraph. In those two weeks the reviewer caught the builder every time, and never the reverse, which says less about the models than about the roles: the author is the wrong reviewer, whichever model wrote the code.

The cost is roughly double the tokens on the tickets you run both on. Against a false cost claim in production, a parity metric that could not fail, and a growth figure that depended on row order, it was the cheapest insurance we bought all quarter.

If you are giving an agent live market data, either one can use it: get a free API key and point it at the MCP connector.