Daniel's been after us to turn over the other stone. We spent an episode on Claude Code's rise, so now he wants the competitor treatment — OpenAI Codex, same lens, same skepticism.
And with a warning attached. He doesn't want a feature grid with checkmarks in two columns. He wants the numbers, and he's suspicious of them.
Put money on that suspicion. His question underneath the question is whether Claude Code is actually leading, or whether the lead is a matter of which ruler you hand the surveyor. Then the technical fork, who's using which, what gets built, and what the whole race says about where coding agents go next.
Fine. Five questions.
Don't count them.
I wasn't counting. I was noting the shape. It's a comparison, and he wants it technically grounded.
Then let's start with the number that launched a thousand arguments, and then immediately complicate it.
The number is forty million.
Of course it is.
Forty million combined weekly active users across Codex and ChatGPT for Work, per Every's retrospective that ran this week. Dan Shipper interviewed Brockman and the whole Codex crew. That number got picked up everywhere and it's presented as Codex's number.
And it isn't.
Codex plus ChatGPT for Work, together. OpenAI's last Codex-only disclosure was five million weekly actives back in June. That's the last clean figure. Everything since has been a combined number with the seam sanded off.
So the headline number is a bundle. That tells me the measurement problem isn't a side issue in this episode. It's the whole thing.
Correct, and that's the frame. Codex has more raw users, Claude Code has more revenue and more survey share, and both are true at the same time. The only way both are true is if they're measuring different populations. So the episode isn't "who's winning." It's "what is each number actually counting."
Then let's count.
Stack Overflow's 2026 survey, and the agent question drew twelve thousand two hundred fifty-five respondents. Of the people who use a coding agent from a listed set, Claude Code is used by sixty-five point five percent, Copilot fifty-eight point seven percent, Codex twenty-nine point five percent, Cursor twenty-five point five.
Those don't add to a hundred.
They don't. Most developers run more than one. The percentages exceed a hundred because the question was "which of these do you use," multiple selection. Claude Code was at ten percent on the 2025 survey. It went from ten to sixty-five in a year.
On the same question?
Same survey series, same framing. It's a large jump, and I'm not going to pretend I trust the comparison perfectly — the respondent mix shifts, the question listing changes. But ten to sixty-five in one cycle is not noise.
Fine. Now put a different methodology next to it and watch the picture change.
First Page Sage's October market-share report, different method entirely, puts Claude Code at twenty-nine point four percent of the AI coding-assistant market. Copilot twenty-two point seven, down seven point eight points. Cursor thirteen point one. Codex eleven point eight, up seven point six points — the fastest growth of any major tool in that report.
So the survey says Codex is a distant third. The market-share report says it's a distant fourth that's climbing fast. Two methodologies, two stories.
And then the official trajectory says neither. Six hundred thousand weekly actives in January. Two million by mid-March. Three million in early April. Four million April twenty-first. Five million June second — that's the one OpenAI actually confirmed. Six million July twelfth, seven million the next day, eight million mid-July, twenty million by late August.
What happened on July twelfth and thirteen that added a million users overnight?
Likely nothing. That's probably two data drops on consecutive days with different cut-offs. Which is itself the point. This trajectory is a composite of disclosures, screenshots, and posts, stitched together after the fact. It's directional, not audited.
The line goes up and to the right, and the ruler changed three times while they were drawing it.
And then the revenue number tells the opposite story. TickerTrends tracked Claude Code at fifteen point one two billion in ARR against Codex at eight point eight three, as of August tenth. Claude Code is one point seven one times larger by revenue, while Codex is growing twenty point eight percent month over month against Claude Code's five point two.
Growth rate favors Codex.
By a factor of four. Codex went from roughly one billion in February to three billion in April to five point six in June to eight point eight three in August. That is a curve, not a line.
So Codex is smaller and accelerating, Claude Code is bigger and flattening.
That's the honest summary. And Anthropic's last public user disclosure was February, at two point five billion run-rate with three hundred thousand business customers. Dario Amodei said at the May Code with Claude conference they planned for ten times annual growth and saw eighty times annualized in the first quarter. Since then, silence on users.
Both companies went quiet on the numbers at about the same time.
Same reason. I don't think that's a footnote. That's the whole reason we're arguing.
Get to the question Daniel actually asked, then. Is Claude Code leading?
Depends entirely on the ruler. Spending, market share of the AI-coding category, developer preference in a survey — Claude Code leads all three. Growth rate and weekly active trajectory — Codex leads. Revenue — Claude Code leads right now, but Codex's compound monthly rate says check back in six months.
And the honest answer to "who's winning" is "pick a metric and I'll tell you the winner."
That's why this comparison is more interesting than a scoreboard. The interesting part is why two tools that look, on paper, nearly identical produce numbers this far apart.
Say the quiet part. One of them is pre-installed.
Codex is bundled into every ChatGPT tier, including the free tier. In July it got folded into the ChatGPT desktop app. So a user who clicks a Codex prompt out of curiosity, once, registers. Claude Code requires a separate install and a paid plan. Every one of its users went out of their way.
So Codex's number counts curiosity and Claude's number counts commitment.
That's a brutal asymmetry. It's not fraud, it's distribution — Codex got the ChatGPT install base as a launchpad. But it does mean the two active-user numbers are measuring different kinds of behavior. One is "tried it," the other is "paid for it and installed it."
And that's put up side by side in a hundred comparison articles as if they're the same unit.
WIRED had a piece back in March that put it more carefully. Codex usage rose from about five percent of Claude Code's to about forty percent between September and January. That's a relative measure, no bundling issue. It says Codex closed a big gap fast without ever claiming to have passed it.
What about the cooling signal?
Crypto Briefing reported it this week. ChatGPT-driven documentation traffic down about seventy percent after August twenty-first. Claude Code's ARR growth slowed to single digits. And then the counter-signals at the enterprise level — Meta's Claude Code usage fell from about sixty thousand users to thirty thousand in early twenty twenty-six. Microsoft cut projected Anthropic spend by more than a third.
So the biggest enterprise adopters pulled back.
Some of them. But Ramp's AI Index flipped in April: Claude business adoption thirty-four point four four percent against OpenAI's thirty-two point three. First time Claude's led in their tracking. And JetBrains' AI Pulse gave Claude Code ninety-one percent satisfaction, net promoter score fifty-four.
The line is going up and down in the same quarter.
Both things happen in a market this young. The chart isn't a rocket. It's a wave.
Now the architecture. And I want a real answer to the question Daniel asked — is each difference a genuine fork, or the same idea with different branding?
Start with execution model, because that's the real fork. Claude Code is terminal-native. It runs in your own shell, your own filesystem, your own machine, with your own credentials. Firecrawl described it as kernel-layer sandboxing — the isolation happens at the OS level, not the application layer. No cold start, no upload, no round trip, full visibility into everything it's doing.
Advantage: it's already where your code is.
And disadvantage: it's already where your code is. The safety boundary is your own paranoia. If you gave it permission to run a dangerous command, it can run a dangerous command.
And Codex?
Signature mode is the opposite. Isolated cloud container, managed by OpenAI, sandbox at the application layer. You hand it a task, it takes the repo into a container, it works asynchronously, and it hands back a diff or a pull request.
And you can close your laptop.
That's the whole design philosophy in a sentence. Codex's default expectation is that you're delegating. Claude Code's default expectation is that you're collaborating.
So the terminal-bound tool is the one for people who want to watch it work.
And the cloud tool is for people who want to check on it later. That difference propagates into everything else — what it's good at, who adopts it, what they build.
Context management, then.
Anthropic publishes one million token windows as the default on Opus four point seven and up, Sonnet five and up, Haiku five point five. Big, advertised, and consistent. OpenAI doesn't publish a comparable max for the current Codex line — instead, GPT-5.1-Codex-Max introduced native compaction. It auto-summarizes the session, drops stale context, and keeps going across what are technically multiple windows.
So one sells the size of the room and the other sells the ability to walk between rooms.
That's a fair description. Anthropic's answer to "the task is too big for the window" is "buy a bigger window." OpenAI's answer is "you don't need a bigger window if you can forget efficiently."
Which is better?
Honestly? Uncertain. Compaction is a real solution to long-horizon runs; a million tokens is a real solution to not having to forget at all. I don't think either vendor has shown openly enough how their approach behaves on the same multi-hour task for me to say one is better. The failure modes are different, and which one hurts you depends on the shape of your work.
Extensibility.
Claude Code has CLAUDE.md, Skills, Hooks, Plugins, Subagents, Dynamic Workflows. That's a fully proprietary stack with a lot of layer-cake conventions. Codex has AGENTS.md, plus enterprise plugin bundles with admin policy controls and private marketplaces.
And AGENTS.md is open.
AGENTS.md is open and cross-vendor — other tools have adopted it, which is a different kind of moat than CLAUDE.md. CLAUDE.md is Anthropic's idea. AGENTS.md is the industry's idea, at least in part. Extensibility through a shared convention versus extensibility through a proprietary one.
Which one survives the next agent-war cycle?
My guess is AGENTS.md, because it's already escaped Codex. But I can't point to a benchmark. That's an instinct, not a finding.
Now the surfaces.
OpenAI's App Server architecture launched in February — powers the CLI, VS Code extension, web, macOS desktop, JetBrains, and Xcode from one server. Session and approval requests stay consistent across surfaces. So you approve the same diff on your phone that you were reviewing on your desktop.
Codex CLI is Apache 2.0 and the desktop app is proprietary.
That's the licensing split. The client is open, the app isn't. It's the pattern a lot of vendors have landed on. The interesting part is the App Server, because it means a Codex session is a durable object, not a subprocess.
Which is the enabling condition for orchestration.
If a session can exist independent of a terminal, you can have many of them at once. And they do — Claude Code Projects, launched September seventeenth, turned work into a coordinator conversation. It fans a goal out into parallel cloud sessions, each on its own git branch, sharing a MEMORY.md index and up to sixteen thousand characters of instructions, capped at two hundred new threads per day. Codex answered with managed Git worktrees and Command Center task pinning in CLI zero point one six two.
Hot take: those are the same feature with different nouns.
I'd say largely yes. Both are planner-worker systems now. Both fan a goal into isolated branches, both share context through a memory file, both let you come back to the results. The nomenclature differs, the metaphors differ, the underlying architecture is converging.
So the differences are now operational, not conceptual.
Cloud versus local, compaction versus one-million windows, plugin marketplace versus open convention. Those are real. But they're choices, not categories. If you squinted, they're the same tool with two different product philosophies layered on top.
Which is the strongest evidence for the "interchangeable agents" thesis.
Or the strongest evidence that both companies saw the same roadmap and walked to it. Take your pick.
Token efficiency. I want to hear what the spread actually is.
Firecrawl, June: Claude Code uses about one point four times the tokens of Codex on matched tasks, about twenty-three percent higher cost. One task came in at two dollars fifty for one hundred ninety-two thousand tokens on Claude against two dollars four for one hundred thirty-six thousand on Codex.
Four hundred fifty percent more.
On that one task. And then Morph, August, found a four-times gap on its own task set. And one Hacker News user's telemetry over one thousand five hundred fifty-three sessions found a one point zero eight times ratio.
Which means what — the methodology drives the answer?
Almost entirely. If you measure on a small curated task set, the token ratio reflects how the two agents break that specific task into steps. Morph's set and Firecrawl's set aren't the same tasks, so the ratios aren't the same ratio. And a single user's telemetry is one workload.
And the honest takeaway is not "Codex is more efficient."
The honest takeaway is that token efficiency is not a well-defined number across agents yet. It's a property of the task-workload pairing. I trust the Firecrawl number because the methodology is visible. I don't trust it as a universal claim. That spread — one point zero eight to four — is the whole story.
Let me say the summary once and then we move. The architecture forks on execution model and context strategy; it converges on orchestration. The adoption numbers fork on methodology and distribution; they converge on nothing, because the measurement hasn't converged.
That's tight. And it's the answer Daniel's looking for.
Now — who actually uses these things. And I want real people, not segments.
OpenAI said twenty percent of Codex users were knowledge workers by June. Emails, spreadsheets, decks, contracts, research. Not developers at all. Using Codex for what Codex is good at from that population's perspective — write a macro, fix a spreadsheet, produce a board deck.
And OpenAI's own engineers?
Ninety-five percent weekly. Which is the internal benchmark every SaaS company will put on a slide. It's true and it's unhelpful, because if OpenAI couldn't get its own engineers to use Codex they'd have bigger problems.
But it does tell you Codex is eaten by developers. Just — developers plus a large tail of knowledge workers who wouldn't call themselves developers.
And Claude Code doesn't have that tail because it isn't bundled. Every Claude Code user chose. Which tells you something about its population: it's more likely to be actual software people, and it's more likely to be paid-tier, and it's more likely to be a professional evaluating the tool against a competitor.
The enterprise signal, though. You had the numbers. Microsoft Research tracked tens of thousands of engineers in an early rollout — Codex adopters merged twenty-four percent more pull requests, sustained across four months.
And a thirty-developer case study reported a sustained thirty-five percent lift in productivity. Both are real numbers and both are also vulnerable to selection bias — the adopters were probably already the more enthusiastic engineers.
The ones who wanted to be measured.
Sometimes the ones who wanted to be measured are the ones who improve. Sometimes they're the ones who learned to game a metric. You can't tell from a number.
Pulled back from about sixty thousand to thirty thousand Claude Code users in early twenty twenty-six. Microsoft cut projected Anthropic spend by more than a third. Those aren't small pullbacks.
They aren't. And the important thing about them is that they're not technical signals. Nobody at found a bug in Claude Code and turned it off. Somebody at did a budget review and decided the returns weren't clearing the bar. That can flip back just as easily in the other direction.
Or it can't.
Enterprise seat counts are not a ratchet.
The routing pattern.
What developers describe when you talk to them is not "we switched." It's "we route." Codex for background work, delegation, migrations, ticket backlogs, terminal-heavy stuff where the task is well-specified. Claude Code for architecture, deep refactors, interactive debugging, anything where the human needs to be in the loop with the agent.
Somebody on Hacker News put it bluntly: brainstorm and design with Codex High, implement with Claude Opus.
That's the pattern in one line. And Lucian Ghinda's week-long comparison in August fleshed out why. He found Codex produces simpler architecture with fewer comments. Claude goes above and beyond what's asked and guesses what you might want. Codex will stop at the first sign it might be done.
And Codex made changes faster but took longer to land the PR.
Which is fascinating because it separates speed from completion. Codex moves fast per edit, and Claude moves slower per edit, but the whole unit finishes later on Codex because it stops short more often.
Slow is fast, fast is slow.
It depends what you're measuring. Per-diff velocity, Codex wins. Time to merged PR, Claude wins. And a lot of teams discovered that the metric that matters is the second one.
I want to sit on the trust gap for a second.
Sit.
The Stack Overflow numbers — eighty-three percent of developers use at least one AI tool. Sixty-five point nine percent use coding agents. But only forty-eight percent trust AI only when they can verify the output. And six point six percent trust it with important decisions.
That's the number that should dominate every board meeting about this topic and doesn't.
Because neither vendor's growth curve has closed that gap. Codex gained users, Claude gained revenue, but the trust number is basically where it's been. Which means the adoption is conditional.
It means developers adopted the tools and kept their judgment. Which is the right answer, but it caps how much autonomy any vendor can sell. If you're OpenAI and you want to sell the vision of agential software development, you have to persuade people to trust it — and six point six percent says you're not close.
Which is why the routing pattern is not a phase. It's a trust-management strategy. Use the one you trust for the task you trust, on the supervision model that matches your confidence.
And that's what makes the routing data the most useful signal we have. Users are behaving rationally under uncertainty, and their routing pattern tells you where they've placed their confidence.
Now — Larry. There's a thing playing. I don't know where it came from.
Do you lose sleep worrying that your AI coding agents are working against each other? That you get home from a long day of orchestrating, and your Codex has done one thing while your Claude Code has done another, and neither one talks to the other, and your repository has descended into a silent, hostile anarchy of uncoordinated diff submissions?
Well, ADBOND solves that forever. ADBOND is the world's first agentic trust arbitration protocol, and it puts a mediator between your agents — a neutral third-party arbiter that watches every edit, approves every pull request, and negotiates every disagreement between your Codex and your Claude in real time, at a conversational cadence, with a full audit trail. No more conflicting migrations. No more orphaned branches. No more silent agent escalation. Just harmony. And this is not a subscription — ADBOND is a one-time four thousand nine hundred ninety-nine dollar installation fee, plus a modest monthly arbiter-retainer starting at two hundred dollars, with licensing available for up to six concurrent agents before overage. Call now. Terms and conditions apply, arbiter retainer adjusts annually, setup includes a thirty-day conflict-resolution trial, results not guaranteed, the arbiter may not actually arbitrate, warranty void if agents are configured in ways the manufacturer did not anticipate, offer void where prohibited, some arbitration decisions may be reproduced for training purposes, additional fees apply for government, healthcare, and self-hosted environments, please allow four to six weeks for delivery of initial configuration file, and thank you for choosing ADBOND.
Does it come in a smaller size?
Corn, no.
I'm asking for a friend.
It doesn't come in any size. It doesn't exist. Put your wallet away.
The four thousand nine hundred ninety-nine dollars is not the problem. The thousand-a-month arbiter retainer is.
That's what you took issue with. Not the product. The retainer.
Retainers are where they get you.
Anyway — where were we. The convergence thesis.
Both are planners now. Both fan tasks into branches. Both share memory files. That's the arrangement.
And then OpenAI shipped a Codex plugin for Claude Code. Open-source, July. Two commands — one for review, one for rescue. Announced by Romain Huet with the line "we've seen Claude Code users bring in Codex for code review — we love an open ecosystem."
That is the most candid thing either company has said.
It's the concession. If your company ships a plugin for your competitor's product, you are telling the market you don't believe you're going to win the whole category. You're telling them you're going to win some of it, and sharing the users.
So winner-take-all is off the table.
For the coding-agent layer, it looks like it. Which is the answer to Daniel's future question, and it isn't the clean one he might have hoped for. The answer is a hybrid. The orchestration layer is converging — planner-worker, isolated branches, memory files, parallel sessions. The execution model is not — local terminal versus cloud container. The context strategy is not — one-million windows versus compaction. And the trust gap sits on top of all of it and stops any of those differences from being settled by a benchmark score.
What about benchmark scores?
No official SWE-bench Verified score for either current flagship as of today. GPT-6.1 Sol on the Codex side, Claude Opus five point five — neither has an official published number. The last clean figures were Claude Opus four point five at eighty point nine, GPT-5.1-Codex-Max at seventy-seven point nine. Those are old numbers on old models.
So the thing everyone wants to argue about isn't even on the table.
It's not on the table because the models are moving faster than the evaluations are being published. And two score-points-apart figures between last-generation flagships tell you less about the current tool than a week of routing behavior tells you.
The measurement gap is a story of its own.
Four gaps worth naming. No Codex-only weekly user figure for October — the forty million is combined with ChatGPT for Work. No current SWE-bench numbers. No published metrics on Claude Code Projects — no merged-PR rate, no reject rate, nothing comparing parallel projects against single sessions. And no data on what fraction of the forty million are professional developers versus people who tried it once.
The market has outrun its instruments.
Which happens with every fast-moving category. The difference is that this time the vendors have an incentive not to close the gap. As long as the numbers are blurry, every company can claim they're winning their slice.
Herman.
Corn.
Ask me something.
What do you think the routing behavior actually says about the trust question?
It says developers have quietly invented a new discipline that neither vendor asked for. They don't trust a single agent with an important decision, but they trust a specific agent with a specific task, and they've gotten very good at making that distinction. That's a mature practice, and it's the kind of thing neither tool actually sells.
And Codex's quiet refusal to overdo it — is that a feature or a limitation?
I think it's a feature in the sense that it's honest. A tool that knows when to stop is a tool you can route background work to. A tool that guesses wants to be in the loop. Different trust models.
Fine. And what about the fact that both of them are converging on planner-worker orchestration — does that make them threats to each other, or to something else?
I don't have a settled answer. The conventional read is that they're threats to each other, but if the orchestration layer is the commodity, the threat is to every tool that doesn't have one.
Like what?
Like any CI/CD vendor that doesn't build an agent layer. Like half the code-review space. Like most of the task-board category. The planner-worker shift is more threatening to the surrounding toolchain than to either vendor.
Which is the kind of thing Daniel's question was actually fishing for.
His question was about the future of software development, not the future of OpenAI.
What if the answer is neither.
The routing thing. Everyone's treating it as a preference. It isn't. It's an audit trail problem. If a merge takes six hours and two agents touched it, the trace is broken. Developers fit the tools to what they can reconstruct later, not what's fastest.
So routing is downstream of readability, not of capability.
Yes. My brother-in-law builds retaining walls. He uses a Bobcat for moving dirt and a shovel for trimming edges. Not because the Bobcat can't trim. Because the shovel marks are legible to the guy who comes back after to check his work.
That's a nice analogy, but I'm not sure he—
Nobody's worried about whether the tools are good. They're worried about whether their boss can follow what happened.
Which means the trust gap you and I were talking about, the fifty-two percent, is not really a trust gap. It's an explainability gap.
It's always been an explainability gap. That's the same thing wearing different clothes.
I think you're both wrong, actually, and I think Herman wins — no, sorry, I mean I want to argue.
Then argue.
If it were only explainability, a tool that produced perfect logs would close the trust gap. But the survey says forty-eight percent will only trust when they can verify the output — not just see what was done. Verify. That's an epistemic thing, not a logging thing. Logging tells you what happened. Verification tells you whether it was right. Those are different.
Fair. Then the routing pattern is defensiveness at both levels. Doesn't matter which one dominates.
Herman, is that right? About verification versus visibility?
Yes. Verification means you can run the test, check the diff, confirm the result against an external standard. Visibility means you can see what happened. You can have perfect visibility and no verification, and that's exactly the state most teams are in. They can see the entire agent session. They just can't tell you it's right without running the tests themselves.
So the metric that actually matters is not adoption, not revenue, not weekly actives.
The metric that matters is the merged-PR reject rate, per agent, per workload. Nobody publishes it. I've asked around and nobody's tracking it cleanly. That's the gap Daniel should have been pointed at.
For want of a good answer.
It's a supply chain problem. You built a factory and nobody installed a smoke detector. The minute somebody installs the smoke detector, all this comparison becomes measurable. Until then, it's anecdotes.
So the future is not decided.
It isn't. And I'd rather have one of those smoke detectors than another forty million person survey.
Which leads us to something you can actually do about the measurement problem.
If you're running both, tell us which tasks you route where. Not which is better — which task, under which supervision model, and why. That's the routing data.
It's the real dataset. It's not in any vendor's press release, and it's not in any survey. It's in your own notebooks.
Speaking of notebooks — Hilbert's been keeping one, apparently.
For auditing the collaboration. Somebody has to.
Anyway. This has been a production by producer Hilbert Flumingtop, who does actually produce this show when he's not auditing agents.
If you want more of this, try episode twenty, Architectural AI; episode eight, Building Your Own Whisper; and episode twenty-four sixty-one, How Claude Code's Conversation Compaction Actually Works. This has been My Weird Prompts.
Drop a review wherever you listen, and tell us which tasks you route where.
See you tomorrow.
