Every time someone asked me which agentic coding tool to pay for, I gave the lazy answer: "depends on your workflow." So one Tuesday morning in May I cleared my calendar, opened two terminals, and ran the same three jobs through both stacks back to back — a research report rendered to PDF, a marketing landing page, and an interactive analytics dashboard. The conclusion I walked away with, and still hold after months of daily driving both: these tools are not converging on one winner. They are specializing in opposite directions, and the developers getting outshipped right now are the ones still treating the choice as an identity question.
I live in Claude Code — it runs my entire Laravel codebase workflow daily, and the Codex plugin sits installed in that same environment. This comparison isn't a drive-by. Here's what the three runs showed, what's changed since May, and the decision rule I'd give a working developer today.
The three jobs and why I picked them
I deliberately chose the work I actually invoice for, not benchmark puzzles.
A structured research report — long-form, cited, with a cover page and table of contents, delivered as a finished PDF. This tests planning, source handling, and the unglamorous last mile of document rendering.
A SaaS landing page — hero, three feature sections, social proof, pricing, footer, Tailwind, no design system specified. This tests taste: the model has to make visual decisions with no spec to hide behind.
An analytics dashboard — charts, filters, state management, realistic mock data. Multiple components with real interactivity; the kind of build I'd quote at half a day of senior front-end time.
Same prompts, same machine, runs triggered in parallel. I tracked wall-clock time, token burn, and how many times I had to intervene.
Job one: the research report — Codex, clearly
Codex decomposed the prompt into a research outline and started populating sections inside the first two minutes. The finished report read like a consulting briefing: short paragraphs, dense bullets, an executive summary, heavy on citations. It also handled the last mile itself — the output went straight through an integrated render into a finished PDF.
Claude Code opened with clarifying questions — audience, citation style, how speculative the predictions section should be. Genuinely useful questions, but they cost me minutes I hadn't budgeted, and the final document ran considerably longer than Codex's with fewer sources: a longform essay where Codex wrote a memo. Then the PDF step needed a Pandoc pipeline and a dependency confirmation from me.
Codex finished faster, burned meaningfully fewer tokens, and delivered a file I could send to a client untouched. If your week is reports, audits, and briefs, this job alone justifies keeping Codex around.
Job two: the landing page — Claude Code, and it wasn't subtle
Codex shipped a working page quickly. Correct structure, all sections present, passable copy — and the visual language of a 2023 template. Centered gradient hero, three-column icon row, generic pricing table. Nothing wrong. Nothing memorable.
Claude Code took longer and kept polishing past "done": asymmetric hero typography, a textured gradient I'd actually keep, alternating feature layouts, a pricing tier lifted with a shadow pulled from the page's own palette rather than a stock accent color. I would ship the Claude output to a client without revision. The Codex output I'd revise for half an hour first.
The cost of that gap was real — roughly double the tokens. Whether that premium is rational depends entirely on whether your buyers pay for visual quality. Mine do, so for UI work the math favors Claude Code even at 2x the burn. If yours don't, Codex's "competent and shipped" is the disciplined choice.
Job three: the dashboard — the run that reframed the whole test
I expected the dashboard to be the longest build of the three. Claude Code finished it in about two minutes.
I rewatched the timer, because a working dashboard — volume chart, engagement panel, search, leaderboard, settings drawer, state persisting across the drawer toggle — arriving in the time it takes to pour coffee broke my expectations. The mechanism, once I looked at the plan it had produced, was auto-delegation: the build got split into a layout shell, a mocked data layer, chart components, and a filter store, with sub-agents working the isolated pieces in parallel while the lead model stitched them together. Codex produced an equivalent-quality dashboard in roughly four times the wall-clock and several times the tokens; its orchestration exists but needs more nudging.
Across the whole suite the totals nearly evened out — the two tools spent similar tokens overall, just inverted per task. Claude spent more on documents and less on UI; Codex the mirror image. The aggregate was a tie. The lived experience was completely different depending on which job was on the screen. That inversion — not any single score — is the actual finding.
What's changed since the May test
A comparison post that pretends time stopped is worthless, so here's the August 2026 state of both stacks, verified rather than vibed:
On the Anthropic side, Claude Opus 5 shipped on July 24, 2026 — a 1M-token context window, priced at $5/$25 per million tokens, and now the default model on Max plans. If you're carrying prompting habits over from the Opus 4.x era, they actively misfire on this model; I wrote up the migration in what changed about prompting Claude Opus 5.
On the OpenAI side, Codex pricing moved to token-usage-aligned billing in early April 2026, and the Pro tier split into 5x and 20x variants days later. Codex remains bundled across ChatGPT plans — Free, the $8 Go tier, $20 Plus, and the Pro tiers — with heavy professional use realistically landing in the $100–200/month range under usage billing. OpenAI has also announced GPT-5.4's retirement inside Codex at the end of August, replaced by newer models. Translation: the "flat subscription, unlimited agent" era is ending on both sides, and per-task token efficiency now shows up directly on your invoice. That's half of why I keep a close eye on burn — the other half is session length, covered in my token management playbook.
The subscription-math version of this comparison changes often enough that I keep it in a separate piece on Codex vs Claude Code subscriptions; check that before you upgrade anything based on a months-old screenshot.
Where each tool earned trust — and where it didn't
Claude Code earned it on UI builds, deep planning, and any job where output quality scales with spend. The planning conversation that opens a session is still the best "senior engineer thinking partner" experience in the category, and auto-delegating sub-agents are an architectural advantage that compounds across a workweek. It did not earn trust on the document last mile — the PDF pipeline still asks me to confirm dependencies more than it should.
Codex earned it on structured documents, end-to-end shipping, and GitHub-mediated work. Tagging Codex on a PR and coming back to a finished line-by-line review in a cloud sandbox is a genuinely great workflow. Its computer-use QA found a broken anchor and a dead mobile hover state in the landing page it had just built — competent first-party verification that Claude Code needs external tooling to match, which is exactly the gap I fill with Playwright wired into Claude Code. It did not earn trust on visual taste. Fine is not what I sell.
The two tools also compose better than they compete: I run the Codex plugin inside Claude Code and use it as an adversarial second opinion on plans and diffs — one model drafts, the other attacks. If that workflow interests you, I've written up the Codex plugin inside Claude Code setup separately.
The decision rule
If you can only run one:
- Your week skews UI, product surfaces, client-facing polish → Claude Code. The visual quality difference converts to revenue, and the sub-agent architecture makes big front-end builds feel small.
- Your week skews research, documentation, PR review, team GitHub flow → Codex. Faster on structured output, better integrated shipping, cheaper per finished document.
- Your week is both — most professional weeks are — run both and stop optimizing the choice. Route jobs, not loyalty. The combined cost of two subscriptions is still the cheapest senior-capacity line item in any budget I manage.
The Tuesday experiment didn't crown a winner, and months later I'm glad it didn't. It gave me a routing table — this job goes to that engine — and the routing table has been worth real hours every week since. Run your own version: pick the three jobs you actually do for money, run them through both stacks back to back, and write down what each tool did well before brand preference edits your memory.
A routing table is a small artifact with a long payback, and the maintenance is the part nobody budgets for — both engines ship changes monthly, so the table has to be re-run, not just built once. Building and maintaining one against a team's real shipping work is on my services page.
