Follow ZDNET: Add us as a preferred source on Google.
ZDNET's key takeaways
- Claude Code dominates, but Codex has crucial advantages.
- Cost, workflow, and trust matter as much as code quality.
- Whichever AI you choose, human review remains essential.
Earlier this week, I explored the differences (or, really, amazing similarities) between agentic coding AIs, ChatGPT Codex and Claude Code. I explained how I use both tools and provided some suggestions for exploring them.
Also: Claude Code vs. Codex: Why I use both and how to choose one (if you must)
But I'm just one guy. In this article, we'll broaden our selection to more than 100 developers who actively use AI coding tools.
To be clear, this isn't a fully scientific survey. I posted a call for developer opinions to HARO and Qwoted, two services that specialize in connecting experts who want to be quoted or interviewed with journalists and analysts who might choose to quote or interview them.
Also: Claude's Record-a-Skill cut my research from hours to 30 minutes - but the magic has limits
Still, 138 people responded to my question, "Do you use Claude Code or OpenAI Codex? If so, which did you choose, and why?" By the way, if you want to know how I aggregated all these responses, here's the story about the Claude skill I used to do the job.
The top-line data point is interesting. Three out of four developers use Claude Code. This mirrors my discussions with developer colleagues. Very few of them talk about Codex, and all of them talk about Claude Code.
A little more than a third of the respondents use Codex. And, explaining why the numbers don't fully add up, 22% use both Claude Code and Codex (I'm in that cohort).
Preference trends: Claude Code over Codex #
Naman Ahuja is a software engineer at Meta. He said, "I use Claude Code extensively at Meta because it fits naturally into my development workflow. I treat it like a peer: brainstorming ideas, shaping plans, and executing. It has improved my productivity and lets me focus on more business problems to solve. I'm happy with it, though I review its output carefully."
Another Claude Code user is Mona Rajhans, a senior software engineering manager at Palo Alto Networks. She said, "I use Claude Code daily and manage a team of engineers who do too. What I notice at both levels: it handles context across large, complex codebases better than anything else we've tried. Code quality has held up. Adoption was faster than any tool rollout I've managed in years."
Saurabh Kumar Suresh Jain is a staff data scientist at Nvidia. He said he prefers Claude Code: "I use it daily on a Python analytics dashboard that pulls EDA farm telemetry -- multi-file refactors, writing tests, reviewing my own merge requests before I push. What sold me is that it reads the whole repo before it edits, instead of guessing. Yes, happy with it."
Also: 7 AI coding techniques I use to ship real, reliable products - fast
Srinivas Chippagiri is a senior member of technical staff at Tableau, a Salesforce company. He said, "I use Claude Code for cloud-native distributed systems and AI infrastructure work: service code, pipelines, and the glue around ML workloads. I chose it for agentic, multi-file reasoning. It navigates a large codebase, runs commands, and iterates toward a working result rather than just autocompleting a line. I'm happy with it. It removes real toil on the routine parts, but the architecture and the final judgment stay mine, and I review everything it touches."
There are some definite preference trends. When it comes to why developers prefer Claude Code over Codex, these were the main themes:
- Claude Code holds a large codebase in context, so changes spanning many files stay coherent (14%).
- Claude Code plans and reasons through changes before making edits, so it needs fewer corrections during complex refactors (9%).
- Claude Code produces better raw code quality and reasoning than Codex on the same task (9%).
- Claude Code works agentically in the terminal against the real repo and real infrastructure, not as autocomplete inside an editor (7%).
- Claude Code got there first, and the skills, integrations, and workflows are now built around it. Switching costs real work (5%).
- Claude Code's agent tooling (skills, subagents, hooks, MCP) supports multi-agent setups that Codex is still copying (4%).
- Claude Code better explains its reasoning as it works, and has more tutorials and community material to learn from (4%).
- Developers trust Anthropic as a company and discussed an active distrust of OpenAI (4%).
Also: Why software factories are back - and how they work in the age of AI
Obviously, these are developers' opinions and reflect their perceptions of the products. We also had single-user responses that shed some light on preferences. These included:
- Claude Code executes money-touching payment and subscription code without guessing, even across many checkout refactors.
- Claude Code checks the math and data with you, not just the code, when working on Bayesian modeling in JAX and NumPyro.
- Claude Code is noticeably better at working from voice dictation.
- Claude Code asks permission before taking destructive actions, providing real auditability and human oversight.
- Claude Code runs on GitHub Actions rather than a laptop, so it keeps working while you're away from the desk.
- Claude Code finds and fixes its own bugs.
Preference trends: Codex over Claude Code #
Interestingly, I didn't get responses from developers who prefer Codex at institutional heavy hitters like Nvidia, Meta, Palo Alto Networks, and Salesforce. That result doesn't mean big enterprises don't choose Codex. It just means that my informal survey didn't generate those responses.
Also: 'Specialists aren't required' anymore: How to stay valuable in an AI agent workplace today
Chris Seymour, founder of GS Consulting, said he prefers Codex: "I prefer OpenAI Codex and use it for secure AI and cybersecurity software, testing, research automation, and deployments. I've been happy with it, particularly when the work crosses code and infrastructure. I also greatly prefer it over Claude as OpenAI provides much more usage for the same price. I never run out of tokens, and I'm using it every day."
Anthony Woo is CFO and co-founder at Torus, a nonprofit providing curbside composting. He said, "I used both Claude Code and OpenAI Codex to rebuild Torus's website. Claude Code was stronger at understanding the broader codebase, but Codex delivered more precise, reliable edits with less cleanup. I favor Codex and am very satisfied with how much it accelerated development."
Adam Dalloul is a computer science student and the founder of EmpirioLabs AI, which is part of the Nvidia Inception program. His perspective is interesting because he hosts 100-plus AI models on one API, and he suggested the harness, not the model, is what he's choosing. He said, "I use both Claude Code and Codex apps, not CLI, as daily drivers, each with their own model. My assessment is that as a harness, Codex is far better. The app is much more stable, clearer, and has better QoL features, context management, and compacting."
Looking at responses overall, there are some definite themes for why some developers prefer Codex:
- Codex provides more usage for the same money, with no weekly limits (7%).
- Codex fits into an existing OpenAI/ChatGPT workflow, reducing context switching between thinking and building (6%).
- Codex is predictable and low-maintenance, landing changes with less cleanup and less babysitting from week to week (5%).
- Codex runs the whole loop autonomously, inspecting the repo, changing files, testing, and then handing back a diff to review (4%).
- Codex offers faster turnaround that keeps developers in flow, while long waits break concentration (2%).
Some of the individual comments included:
- Codex offers a better harness, with a more stable app, clearer UI, and better context management and compaction.
- Codex handles work that crosses code and infrastructure cleanly.
- Codex's image generation capabilities have improved sharply alongside its coding capabilities.
- Codex was the best model tested for code reviewing, catching edge cases that the primary author missed.
Developers who use both #
One in five developers said they used both Claude Code and Codex. In my comparison article, I explained that I use each tool on its own project.
However, I recently completed an emergency mitigation project, and I used both tools together. I was very impressed with how successful that deeply urgent weekend push turned out to be. Other developers share my opinion.
Michael Rangel is the founder of finance firm Novo. Novo has raised over $170M of venture capital. Rangel said the firm currently services over 250,000 small businesses across all 50 states. He said, "I use both. Claude Code is my hands-on-desk tool. Codex powers my Hermes agent on the go. Pitting them against each other, with one building and the other reviewing, yields better results than either model alone and gives me, as a non-technical founder, more confidence in the work."
Wesley Cable is the founder of a marketing operations agency for local service businesses. He reported, "I use both. Claude Code runs my whole agency day-to-day, and I keep Codex on tap as a second opinion when a build gets stuck. Codex wins on price and limits. Claude wins on judgment. For a one-person business, judgment is worth more."
Also: 'I'm not a programmer' anymore: Linus Torvalds on the only two tools he uses now
Kai Blakeborough is an enterprise AI strategy and governance advisor. He told us, "Both. At $20/month each, the value is absurd. Claude Code is my visioning and planning tool. Codex handles execution and review. Instead of being locked into one model or harness, I get the best of each, and both improve each other's outputs. I couldn't be happier!"
Finally, Rory Bokser is a Montreal-based developer who is all in with both AIs. He wrote, "I use both on their maximum plans and exhaust all of the weekly usage consistently. I find Claude Code (Fable in particular) to win on creative, more human-level problems, while Codex is more efficient on cost per task. Claude Code is my planner and brainstormer, and Codex is my workhorse. My typical flow for hard problems is planning with Fable, and handing over to Codex for implementation. I also use both as adversarial reviewers for all pull requests."
Overall sentiment trends #
Beyond choosing one tool or another, our pool of developers shared additional insights about their use of AI coding tools:
- 22 respondents told us they like to run the two tools against each other. One writes and the other reviews or audits before anything ships.
- 18 respondents told us they are non-engineers who are using these tools to ship production software. The respondents included founders, marketers, operators, an artist, a photographer, a urologist, and a plumber.
- 14 respondents said human review is non-negotiable. Every AI-authored change is gated like a junior developer's pull request.
- 14 respondents use these tools to run their businesses rather than write software. Their uses include automations, dashboards, and back-office operations.
- 14 respondents said cost and usage limits are their biggest complaints. They cited weekly caps, token burn, and spending that is hard to track until the money is gone.
- 11 respondents use multi-agent setups, in which an orchestrator directs specialized subagents.
- 8 respondents treat the tool as a colleague, peer, junior developer, or new hire rather than as software. This is my approach, as well.
- 8 respondents said a terminal-first workflow is part of the appeal rather than an obstacle.
- 6 respondents said the setup matters more than the model. The harness, context files, permissions, and constraints can determine the outcome.
- 6 respondents warned about silent failures. The code can look right and run while still being quietly wrong. Oh, yeah. So very, very true.
- 6 respondents selected their tool based on trust and company values rather than benchmarks.
- 5 respondents switched tools and stayed. Four moved to Claude Code, and one moved to Codex.
What do you think? #
This article represents just a small slice of the millions of developers who have embraced agentic coding tools. Because they self-selected into responding, they may not be as representative as a study would have been if we had been able to include a true cross-section of developers worldwide.
Even so, the insights shared by these coders are valuable. I identify with much of what the developers reported in their answers. How would you answer the questions we asked our respondents? Let us know in the comments below.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.