For Claude Code users who joined for the coding and stayed for everything else.
If you adopted Claude Code, you probably did it for a reason that still holds: it produces excellent code, reasons well over long context, and keeps you in the loop. Most experienced builders who've tested both still rate Claude Code's output cleaner in blind reviews. None of that has changed.
But the more interesting story in 2026 isn't which tool writes better code. It's that the two products have been quietly evolving along different philosophies — and OpenAI's bet is starting to pay off for a group Claude Code wasn't originally designed to serve: people who aren't professional developers.
The numbers make the shift hard to ignore. In its June 2, 2026 report The Next Era of Knowledge Work, OpenAI disclosed that Codex had passed 5 million weekly active users — more than 6x its level when the desktop app launched in February. More telling than the headline number: knowledge workers now make up roughly 20% of those users and are adopting Codex more than three times as fast as developers (figure vendor-disclosed and confirmed by Axios). Among that cohort, the fastest-growing tasks week-over-week are data analysis (+110%), research (+37%), and the creation of knowledge artifacts — reports, memos, PDFs, spreadsheets — at +36%, with PDF and spreadsheet work specifically up more than 50%.
A caveat worth stating plainly: these are week-over-week growth rates from a report OpenAI released partly to reclaim narrative momentum against Claude Cowork ahead of both companies' IPOs, so treat them as directional. And "3x faster" is a growth rate, not absolute volume — the majority of those 5 million users are still developers. But the direction is unambiguous, and it points at a question worth asking: if you're a non-developer, is the tool you picked still the right fit?
The architectural fork
The two tools made opposite bets, and the bet is the whole story. Claude Code is terminal-native and local-first. It lives in your shell, reads your codebase, shows its reasoning as it works, and asks before doing anything risky — supervised pair programming. For developers, that transparency is a genuine strength.
Codex bet the other direction: a cloud-based agent that runs autonomously in an isolated sandbox, accessed through a unified app (web, desktop, CLI, IDE). You hand it a task; it spins up an environment, does the work, and hands you the result — delegation rather than pairing. For a developer, Claude Code's visibility is a feature. For a non-developer, it's friction: reading shell output and interpreting error traces requires background most knowledge workers don't have and shouldn't need. Codex abstracts that layer away. That single design difference is why the experience gap is so large for non-technical users, even though both tools are highly capable.
What changed: Codex stopped being just a coding tool
OpenAI has spent 2026 pushing Codex well past code. The April desktop update turned it into something closer to a general-purpose work agent: computer use and a built-in browser (it can control the machine and operate live web environments); built-in image generation (GPT Image 2, released April 21, 2026, in the pipeline); rich file handling (files open in a sidebar with previews for PDFs, spreadsheets, slides, and documents); and a proactive mode that pulls in Google Docs comments, Slack threads, and Notion pages to suggest where to start the day. Put together, this is a different value proposition than "AI that writes code." It's an agent that makes things a knowledge worker actually ships — a slide, a spreadsheet model, a one-page HTML site, an image, a document — and increasingly, the ability to hand those outputs to a coworker directly.
Claude Code treats the code as the product. Codex increasingly treats the output as the product. For most non-developers, the output is all they ever wanted.
Where Codex is catching up — and where it isn't
Codex is closing the gap on the office-document work knowledge workers live in: spreadsheets, financial models, decks, and Word-style documents are increasingly first-class, with native previews and image generation in the same surface. What hasn't changed: on pure code quality and long-context reasoning, Claude Code still tends to win blind reviews. The honest framing is that these tools are moving toward each other, but their foundations stay put — one built around your local environment, one around cloud isolation and a unified app. Six months from now the feature lists may look similar; the philosophies won't.
The bottom line for non-developers
If you're a developer who wants control and the best code quality, Claude Code remains a strong default. But if you're a non-technical builder — or you're standardizing tools for a team that's mostly non-technical — the calculus has shifted. Codex now abstracts away the parts that made agentic tools intimidating, while adding the image, document, slide, and sharing capabilities that map to how non-developers actually work. It's worth a second look. Not because it replaces Claude Code, but because for a meaningful slice of users, it may simply fit the way they work better.
Sources: OpenAI, "The Next Era of Knowledge Work" (June 2, 2026); Axios; Constellation Research; Help Net Security. Growth figures are week-over-week and vendor-disclosed — directional, and released amid competitive positioning against Claude Cowork ahead of both companies' IPOs. The "3x faster than developers" figure refers to adoption growth rate, not absolute user volume.
The model doesn't learn from you. The system wrapped around it increasingly does. Holding both at once is the whole point.
A colleague stopped me in the hallway last week with a question I now get almost weekly: "How do I get ChatGPT to remember how I work? How do I train it on my habits so it gets better the more I use it?" It is the most natural assumption in the world. You spend hours with a tool; it should grow into the job the way a new analyst does. The honest answer is in two parts, and they pull in opposite directions.
The hard truth first: the model doesn't learn
A large language model does not learn from working with you. It cannot. The model you are typing to was trained once, at enormous expense, and then frozen. When you ask it something, it is not reflecting on your past conversations or reinforcing what worked last time — it is predicting the next likely word based on patterns baked in during that original training. Your thousand conversations do not accumulate into a smarter model. The experience of using these tools feels like a relationship; it is not. Every session, the underlying model wakes up with the same fixed knowledge it had the day it shipped.
The softer truth: the system remembers
Here is the other half. Most leading AI products now bolt a memory onto that frozen model. ChatGPT, Claude, and others keep a running file on you, drawn from your chat history, and quietly slip the relevant bits back into the conversation before the model answers — your role, your industry, that you prefer prose over bullet points, that you write for executives. None of that lives in the model. It lives in a side file handed to the model each time, like a briefing note passed to a consultant before a meeting.
This is why ChatGPT can feel like it knows you, and why it's so hard to leave. The switching cost isn't the model, which is roughly comparable across vendors. It's the accumulated context — walking to a competitor means starting over as a stranger. The drawback is that this memory is mechanical: it records your stated preferences and retrieves them. It does not develop judgment about you. It is a filing cabinet, not a seasoned colleague.
The notebook and the expertise
The cleanest way to hold this distinction is to separate a notebook from expertise. When a new analyst joins, they keep a notebook — your preferences, the format you like, the clients to never cc. Useful, external, portable — and not the same thing as the analyst getting better. The deeper change you actually want is expertise: judgment baked into the person through repetition until it is no longer something they look up but something they simply are. Today's AI memory is the notebook. The model's knowledge is the expertise, and it was finalized before you ever met it. Most companies buying "AI that learns" are buying a very good notebook and being told it is expertise.
Enter "dreaming"
The frontier is the attempt to close that gap, and in the last month both major labs shipped features they literally call dreaming — borrowed from the way a brain replays the day during sleep. The name papers over the fact that the two companies built quite different things.
OpenAI's dreaming is a smarter notebook. An early version arrived in April 2025; what rolled out in early June is a rebuilt, far more capable memory architecture, reaching paying U.S. users first and free accounts over following weeks. It synthesizes your memory across many conversations without you asking, and keeps it current — revising "I'm traveling to Singapore in July" into "I went to Singapore in July 2026" once the trip is over. It could finally reach the free tier because OpenAI cut the compute to run it roughly fivefold. The accuracy gains are striking but worth caution — they're OpenAI's own internal evaluations. And a quieter finding for anyone in a regulated shop: an outside study of ChatGPT memory entries found the overwhelming majority were created by the system on its own initiative, not at the user's explicit instruction. The notebook writes itself, and you may not always know what's in it.
Anthropic's dreaming, shipped a month earlier for its agent platform, aims higher. It's a scheduled process that reviews an agent's past work sessions, spots recurring mistakes and the workflows it keeps converging on, and rewrites its own memory to be sharper next time. In one early pilot, a legal-AI company reported task-completion rates climbing several-fold once agents could carry concrete lessons between sessions — down to specific file-format quirks and tool workarounds. This is closer to the notebook learning to take better notes. But — and this matters — neither one changes the model's weights. Both are still notebooks, however clever. Dreaming is the notebook getting organized, not the analyst getting wiser.
Where it is actually heading
The genuinely interesting work tries to turn experience into expertise — to write lessons back into the model itself rather than a side file. The literature even names the wall current approaches hit: "context collapse," where a notebook gets so stuffed with accumulated experience it stops being useful. The proposed escape is to internalize experience into the model's parameters. This is no longer fringe: Yuandong Tian, who spent a decade leading reasoning research at Meta, recently left to co-found Recursive Superintelligence, which emerged from stealth at a reported multi-billion-dollar valuation to pursue exactly this. Whether that's two years away or ten is genuinely contested. The direction of travel is not.
What to actually do with this
First, stop waiting for the tools to learn your business by osmosis. They will not. Capturing institutional knowledge has to be a deliberate act — writing the notebook on purpose, not hoping the model absorbs it. Second, treat memory as the sticky asset it is. The vendor that holds the richest context on your people is the one hardest to leave; decide on purpose whether that lock-in is one you want. Third, when a vendor sells you "self-improving AI," ask the only question that matters: is this changing the model, or just the notebook? Today, almost always, it's the notebook. That's not a reason to dismiss it — a well-kept notebook is worth a great deal. It's a reason not to mistake it for the thing it's dressed up to be.
Sources: OpenAI and Anthropic product announcements (memory / "dreaming" features, April 2025–June 2026); independent study of ChatGPT memory-entry provenance; academic literature on "context collapse" and parameter-internalized experience; reporting on Recursive Superintelligence (Yuandong Tian). Accuracy/adoption figures are vendor-reported internal evaluations — directional.
On the evening of Friday, June 12, Anthropic abruptly disabled its two most powerful models — Claude Fable 5 and Claude Mythos 5 — for all customers worldwide, with the shutdown carrying into early Saturday. It appears to be the first time the U.S. government has forced a publicly deployed frontier AI model offline — and it's the disruption that delayed this issue.
What happened. Anthropic says it received an export-control directive at 5:21pm ET Friday, citing national-security authorities; a U.S. official confirmed the Commerce Department sent the letter. The order prohibits access to both models by any foreign national — inside or outside the U.S., including Anthropic's own foreign-national employees. Unable to reliably distinguish foreign nationals from domestic users in real time, the only compliant option was to shut both models off for everyone. AWS revoked access on Amazon Bedrock the same day. Both models had launched just three days earlier, on June 9; Fable 5 was Anthropic's first general release of a Mythos-class model (a tier it positions above Opus), making the takedown an abrupt reversal of a major launch.
The dispute. The government's stated concern is a reported method of "jailbreaking" Fable 5. Anthropic says the directive arrived without specific technical detail and largely verbally; it reviewed a demonstration and characterized the findings as a small number of previously known, minor issues — not a novel or serious vulnerability. White House AI adviser David Sacks said the administration issued the export control reluctantly after Anthropic allegedly declined to patch the jailbreak first, and that the restriction should lift once it's fixed — "the ball is in Anthropic's court." Reporting (Semafor) indicates Amazon flagged the jailbreak, with CEO Andy Jassy in contact with the administration. Anthropic is complying but pushed back hard publicly, arguing a narrow potential jailbreak shouldn't recall a commercial model deployed to hundreds of millions, and that applying this standard industry-wide would "essentially halt all new model deployments for all frontier model providers."
What's unaffected. All other Claude models — including Claude Opus 4.8 — remain fully available. (Fable 5 and Mythos 5 already carried a mandatory 30-day data-retention designation as Covered Models, a control predating the shutdown.)
Why it matters for us. This is a regulatory-precedent story more than a product story: frontier-model access can be revoked by executive action on short notice with limited technical disclosure — a real planning variable for any enterprise standardizing on a single frontier model or vendor. It reinforces the case for buy-vs-build optionality and vendor-resilience thinking. For now, our day-to-day tooling is unaffected; Opus 4.8 and the rest of the lineup remain available.
Sources: Anthropic official statement ("Statement on the US government directive to suspend access to Fable 5 and Mythos 5"); Bloomberg; TIME; Engadget; The Next Web; The New Stack; TechCrunch; Tom's Hardware (David Sacks remarks); Semafor; AWS Bedrock status notice. Reporting as of June 13, 2026; situation developing.