Claude Code vs Codex: Complete 2026 Guide to 3 Misleading Numbers

14 min read

Search claude code vs codex and you will be handed numbers. A SWE-bench score. A speed multiple. A token count. They look like the kind of evidence that settles an argument, and they get repeated across the pages ranking for this term as though they do.

We tried to trace three of the most-repeated figures back to their sources. One does not appear on the leaderboard it is credited to. One comes from a test in which neither tool ran its own model. One appears in two places with two incompatible accounts of where it came from. Here is what checking actually turned up, and what genuinely separates these tools in September 2026.

First, the Framing Almost Every Comparison Still Uses Is Obsolete

The standard opening is that Codex runs your task in an OpenAI-managed cloud container while Claude Code runs locally in your terminal against your real files. It was a fair description once. Both vendors’ current documentation contradicts it.

Anthropic’s overview page describes Claude Code as “an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools. Available in your terminal, IDE, desktop app, and browser.” It states plainly that Claude Code “runs on several surfaces: the terminal, IDE extensions, a desktop app, and the web”, and that each surface connects to the same engine, so your project configuration follows you between them.

OpenAI’s Codex documentation, meanwhile, documents both local and cloud execution — naming “Local environments” alongside a “Cloud environment,” with the interface exposing a choice between running on your computer and running in Codex cloud. Codex also ships a CLI, an IDE extension for VS Code and JetBrains, a web interface, an SDK, and GitHub and Slack integrations.

Line the two up and they are close to symmetrical: CLI, IDE extension, desktop presence, web interface, SDK, CI and chat integrations, and the option to run locally or in the cloud on both sides. Choosing between them on where the code executes is choosing on a distinction that has quietly stopped being true.

Misleading Number 1: The SWE-bench Score That Is Not on the Leaderboard

The most authoritative-sounding figure in this comparison circulates roughly like this: Anthropic publishes SWE-bench Pro results showing Claude Opus 4.7 at 64.3% against 58.6% for the OpenAI side. It appears in an article that ranks for this term, stated as fact, with no link.

We opened the official SWE-bench Pro leaderboard and checked. Four things are wrong with that sentence.

• Neither score is there. 64.3 does not appear on the leaderboard. Neither does 58.6.
• The model is not there. There is no “Claude Opus 4.7” entry. The Claude Opus entries are claude-opus-4-6 (thinking) at 51.90 and claude-opus-4-5-20251101 at 45.89.
• 64.3% would be state of the art. The top score on the leaderboard is Muse Spark 1.1 at 61.50, with gpt-5.4 (xHigh) second at 59.10. A claimed 64.3% is not just unsourced, it is above the field.
• The publisher is wrong. SWE-bench Pro is published by Scale AI through Scale Labs, not by Anthropic. A benchmark a vendor publishes about itself and an independent leaderboard are different kinds of evidence, and the claim quietly swaps one for the other.

There is a fifth problem that dates the whole claim. Opus 4.7 is not in Anthropic’s current lineup: its September 2026 model overview lists Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5. Opus 4.7 is still an active, supported model — Anthropic’s deprecation page gives it a retirement date no sooner than April 2027 — but the Opus sequence has run 4.7, then 4.8, then 5, so an article presenting it as the thing to compare is two releases behind.

Misleading Number 2: The Speed Test Where Neither Tool Ran Its Own Model

Ask Google which is faster and, when we checked on 6 September 2026, the AI Overview said Codex wins on speed and token efficiency — though AI Overviews are generated per query, so yours may differ. Ranking below it on the same page was a comparison, published in August 2026 and based on a substantial testing effort, reporting the opposite: “Claude Code finished in about half the median time and used 90 fewer tool calls. Codex took longer and made more calls, but the entire run cost around 58% less.”

Both cannot be the headline answer. And the detail that resolves it is in the methodology rather than the result.

That test ran 30 identical multi-app tasks with a 900-second limit, and both harnesses were running DeepSeek V4 Flash — a third-party model that is neither Anthropic’s nor OpenAI’s. Both scored 16 of 30. Claude Code posted a median of 122.7 seconds and 358 tool calls at $0.195 per success; Codex posted 245.0 seconds and 448 tool calls at $0.081 per success.

Read carefully, that is a well-constructed experiment measuring something real: how efficiently each harness drives the same model. It is a harness benchmark. It is not a measurement of Claude against GPT, which is what a reader searching this term almost certainly wants to know, and it is not what the headline number gets used to prove.

It is also worth naming that the publisher sells tooling infrastructure to users of both, so it is not a wholly disinterested party — which does not make the test wrong, and the methodology is more transparent than most in this category.

The most-cited speed comparison between these two tools was run with neither company’s model in either seat.

Misleading Number 3: The Token Figures Nobody Can Source

A pair of token counts circulates in this comparison — roughly 1.5 million tokens for one tool against 6.2 million for the other on the same task. The figures themselves may well be real. The provenance is the problem.

One site publishing them, in April 2026, presents them as the result of its own internal test. A second site, three days later, publishes the same pair of numbers and attributes them to two entirely different companies. Same figures, two incompatible accounts of where they came from, neither with a reproducible methodology attached. Both publishers sell competing AI development products.

We are not asserting that either is dishonest, and we are deliberately not repeating the numbers as data. We are pointing at the pattern, because it is the same one that turns up every time a figure in this category is traced: a number gets a life of its own, and by the third or fourth retelling nobody can say who measured it.

We found the identical failure when tracing the viral token-efficiency stat in our Cursor vs Claude Code guide, where a widely-cited “independent benchmark” turned out to link to a page that did not contain the numbers at all.

AGENTS.md vs CLAUDE.md: Half Right, and Misleading

One structural claim in circulation is that Codex uses the open, cross-tool AGENTS.md convention while Claude Code locks you into a proprietary CLAUDE.md system. The first half is accurate. The second oversells the lock-in.

Anthropic’s documentation is blunt about the default: “Claude Code reads CLAUDE.md, not AGENTS.md.” But the same page documents three one-line bridges. You can import the file — a CLAUDE.md containing @AGENTS.md loads it at session start. You can symlink it. Or you can run /import, which copies a supported agent’s configuration across — instruction files “such as AGENTS.md“, along with MCP servers, commands, subagents and skills. Note the difference: the @ import and the symlink track the file as it changes, while /import takes a one-time copy that goes stale when AGENTS.md is edited.

The claim also conflates two things Anthropic documents as separate systems. CLAUDE.md is written by you and holds instructions; auto memory is written by Claude and holds learnings. Describing “the proprietary CLAUDE.md auto-memory system” merges a file you author with a mechanism that authors itself.

The honest version: if your repo standardises on AGENTS.md, Codex reads it natively and Claude Code reads it after a symlink or a one-line import. That is a real difference in friction. It is not a wall.

What Claude Code vs Codex Actually Costs

Pricing is the part of this comparison that is genuinely checkable, and the structures have converged more than the commentary suggests. Both vendors bundle the coding agent into a chat subscription rather than selling it separately.

Two things stand out. Codex has a genuine cheaper rung at $8 that Claude does not match. And the multiplier language at the top is nearly identical on both sides — 5x and 20x above the standard tier — which suggests convergent thinking about what heavy users will pay rather than a meaningful difference.

What neither vendor gives you is a hard number you can budget against. Both publish usage as multiples and windows rather than fixed token quotas, so any comparison promising you a precise “X times more compute for your $20” is reconstructing something the vendors themselves do not publish. We looked for the source of one such widely-shared figure and could not find one.

What Developers Who Use Both Actually Report

The first-hand accounts are more useful than the benchmarks, and more mixed. One caveat first: the richest discussion thread on this comparison dates from late 2025, so treat it as a snapshot of that moment rather than today.

In it, a developer posting as veidr described Codex as actively counterproductive in his workflow, and was specific about why: “Claude just misses stuff (fine; expected), while Codex comes up with plausible edge-case database query concurrency bugs” that cost him half an hour of investigation before he concluded they were not real. His verdict came with a hedge worth keeping attached to it: “So for me, head-to-head, Claude murders Codex — and yet I know that isn’t true for everybody, so it’s weird.”

He does not write Codex off, though, and the exception he carves out is the most interesting thing in the thread: “What I do like Codex for is reviewing Claude’s work (and of course I have all of them review my own work, why not?).” Even there he notes Codex “sometimes flags nonexistent bugs” — but says that matters less in review, because he can let the two tools argue it out by writing tests.

Another developer in the thread, posting as willaaam, was also on the Claude side, though about models inside GitHub Copilot rather than the standalone agents: “It’s not that Codex is bad, it just feels tonedeaf to me. It writes code in its own preferred style and doesn’t adjust to the context of the codebase.”

Be careful how much weight you put on this. It is one thread, and the accounts in it lean toward Claude Code rather than splitting evenly — which is worth saying plainly rather than smoothing into a false balance. The write-and-Claude, review-and-Codex split is one developer’s practice, not a consensus. The author of the post the thread was discussing, who sells AI workshops and so is not a neutral party, landed on a framing worth borrowing anyway: “The tool you choose should match how you work, not the other way around.”

How to Actually Choose

• Your repo already has an AGENTS.md: Codex reads it as-is. Claude Code needs a symlink or a @AGENTS.md import, which takes a minute and then tracks the same file.
• You are price-sensitive at the bottom: Codex’s $8 Go tier has no Claude equivalent. At $20 the two are level on sticker price.
• You want the agent to review rather than write: at least one developer who otherwise strongly prefers Claude Code keeps Codex specifically for reviewing Claude’s output. That is one account, not a pattern, but it is a cheap experiment to run yourself.
• You are choosing on a benchmark you read somewhere: check what model each side was running before you weight it. In the most-cited recent test, the answer was “the same third-party one.”
• You cannot decide: both bundle into a $20 chat subscription you may already be paying for. Running both for a fortnight costs less than the time spent reading comparisons, this one included.

For the broader landscape, our AI coding tools hub sorts this space into the three categories that actually differ, and our Claude Code Review guide covers the review feature that shares a name with the general comparison. The SWE-bench Pro leaderboard we checked is published openly by Scale AI.

Claude Code vs Codex FAQ

The Honest Summary

The claude code vs codex comparison is unusually well supplied with numbers and unusually short of numbers that survive a check. A leaderboard score that is not on the leaderboard, a speed result from a test neither vendor’s model took part in, and a token figure nobody can source are three examples, and they were the first three we pulled. Only the first is outright falsified. The second measures something real but not the thing it gets quoted for, and the third simply cannot be traced — which is a lesser charge, and worth keeping distinct.

Meanwhile the structural claim most guides open with — cloud versus terminal — has been overtaken by both products shipping nearly the same set of surfaces. What is left is genuinely close: similar pricing, similar reach, and a task-by-task preference that differs between honest practitioners. Run both against your own work for a fortnight. It is the only benchmark whose methodology you can see.