75% of Developers Prefer Claude Code Over Codex: What the Data Actually Says

Claude Code is Anthropic's agentic coding tool that runs in the terminal and operates directly on your local codebase. According to a ZDNET survey of 138 developers published in August 2026, approximately 75% use Claude Code as their primary AI coding tool — compared to roughly one-third who use Codex — with 22% using both. The gap is real, and the reasons behind it are worth understanding.

What the Survey Actually Found

The ZDNET survey is the clearest head-to-head data we have right now. A few things stand out:

75% of respondents use Claude Code as part of their regular workflow
Roughly one-third use Codex — OpenAI's coding agent
22% use both, which means the populations overlap but Claude Code has broader reach

One methodological note: 138 developers is a real sample but not a massive one. Self-selected surveys skew toward early adopters who are already experimenting with multiple tools. That said, the gap is large enough that even with sampling bias, it's pointing at something real in the market.

The survey was published the same week OpenAI announced Codex gaining access to local codebases and terminal through the ChatGPT desktop Linux app — a capability Claude Code has had. The timing matters because Codex is catching up on the interface side, which makes the preference gap more interesting, not less.

Why Developers Choose Claude Code

The survey respondents cited specific reasons for their preference. These line up with what practitioners in multi-agent and enterprise contexts have been reporting independently:

Instruction-following at scale. Claude Code handles long, complex task chains without drifting from the original intent. When you're running agents that need to execute 20+ steps across a codebase, losing context halfway through is a hard failure. Claude's extended context window and instruction adherence reduce that failure mode.

Agentic behavior out of the box. Claude Code is built to plan, act, and iterate — not just autocomplete. It can read files, run commands, edit code, and check its own output in a loop. This is the baseline requirement for any serious multi-agent workflow.

Trust boundaries and approval workflows. For teams working in safety-critical or regulated environments, Claude Code's permission model matters. You can configure which tools it can use, require human approval before executing certain actions, and audit what it did. This is not a minor feature — it's the difference between an agent you can deploy and one you can't.

Transparency on reasoning. Developers working on complex systems need to understand why an agent made a decision, not just what it did. Claude Code's outputs tend to include more explicit reasoning, which makes debugging faster and makes it easier to catch errors before they compound.

Claude Code vs. Codex: Side-by-Side

Capability	Claude Code	Codex
Local codebase access	Yes	Recent addition via desktop app
Terminal access	Yes	Recent addition via desktop app
Long-context instruction following	Strong	Improving
Agentic planning loops	Built-in	Available
Approval/permission workflows	Configurable	Available
Multi-agent orchestration support	Yes	Partial
Linux support	Yes	Recently added

The table reflects the state of things as of mid-August 2026. Codex is adding capabilities — local codebase and terminal access are now live on Linux — so the gap on raw features is narrowing. The preference gap likely comes down to how these features perform in practice, particularly for complex, multi-step tasks.

Where This Gets Interesting: Multi-Agent Orchestration

The 75% figure becomes more meaningful when you look at what developers are actually building. Single-file completions are table stakes. The real use case driving adoption is multi-agent orchestration — running Claude Code as an orchestrator or sub-agent inside a larger pipeline.

A typical pattern looks like this:

An orchestrator agent receives a high-level task (e.g., "refactor the authentication module and update all downstream tests")
It breaks the task into subtasks and delegates them to specialized Claude Code instances
Each sub-agent operates on its slice of the codebase with specific tool permissions
Results feed back to the orchestrator for review before anything gets committed

This kind of workflow requires an agent that can follow constrained instructions reliably, operate within defined tool scopes, and produce outputs that are auditable. Claude Code's design — particularly around tool use approval and context retention — fits this pattern well.

For teams building on top of API gateway patterns, Claude Code connects cleanly to external systems through the MCP (Model Context Protocol) ecosystem. You can wire it to internal APIs, CI/CD pipelines, or monitoring tools and have it act as an active participant in your engineering workflow rather than a passive autocomplete layer.

What This Means for Teams Evaluating AI Coding Tools

If you're deciding between Claude Code and Codex — or deciding whether to commit to one — here's the practical read:

Go with Claude Code if:

Your workflows involve long task chains or multi-agent coordination
You need configurable approval gates for tool use
You're building in a regulated or safety-critical environment
You need to audit agent decisions after the fact

Codex makes sense if:

You're already deep in the OpenAI ecosystem
You want tight integration with ChatGPT workflows
Your tasks are shorter and more self-contained

Use both if:

You're evaluating which fits your team's actual patterns (22% of the survey respondents do this)
You want redundancy across providers
Different teams or projects have different requirements

The 22% who use both aren't confused — they're being practical. Provider lock-in on AI tools is a real risk, and maintaining familiarity with more than one tool is reasonable engineering practice.

The Governance Angle

One thing the survey data doesn't capture directly but practitioners keep raising: governance. As Claude Code and tools like it move into enterprise workflows, the question isn't just "does it write good code?" — it's "can we control what it does, log what it did, and explain its decisions to non-technical stakeholders?"

Claude Code's permission model and audit trail capabilities are genuinely useful here. Teams building vibe coding workflows — where developers describe intent and the agent implements — need guardrails that don't kill velocity. The sweet spot is configuring approval workflows for high-risk actions (writing to production configs, deleting files, calling external APIs) while letting the agent move freely on lower-risk tasks.

If you're thinking through how to structure those boundaries, the patterns around MCP server configuration and tool scope restriction are worth studying. That's where the real operational work is, not just in picking a tool.

The Bottom Line

The 75% preference for Claude Code in the ZDNET survey isn't surprising if you've been watching how the tool performs in complex, multi-step scenarios. The reasons developers give — instruction following, agentic behavior, control mechanisms — are the right reasons. They reflect actual workflow requirements, not hype.

Codex is closing the gap on surface features. Whether it closes the gap on the things that matter for orchestration and governance is the more important question. For now, the data says Claude Code is where most developers are doing their serious work.