Claire Vo turned a smart lightbulb into a toolbar app.
Not a script. Not a home-automation integration. She told Codex to build an always-on Mac utility so she could manage the bulb from her menu bar, and it did. "THE FUTURE IS NOW," she posted, and for once the caption undersold it.
The same week, Nick Baumann sat in front of a hotkey, hit it, and watched an orb read his screen through Appshots. One spoken command forked into three parallel threads: book a flight, file an expense report, check the calendar. He never opened an app. He just kept talking.
This is the capability leap in its natural habitat. Codex no longer answers questions about your computer. It operates the computer—seeing, clicking, and typing with its own cursor while you use other apps, per OpenAI's own April release. Multiple agents, running on your Mac in parallel, without stepping on your work.
And here is the thing most people watching this get wrong: none of that is what users are complaining about.
The complaint is that it feels slow.
That gap is now the whole story.
Consider the tension in its purest form. Codex has, by any reasonable measure, reportedly won the capability war for general computer operation. It has voice through what Codex Knowledge Base reporting describes as the full-duplex GPT-Live-1 model, which per Codex Knowledge Base reporting listens and speaks simultaneously and decides on its own whether to keep listening or pause. It has screen-reading via Appshots. It has thread-forking, where a complicated task spawns a new branch or pulls context from a previous conversation without you organizing anything first. It has Heartbeats—always-on agents that ping you when a suspicious charge appears or a package is nearby. It reportedly reviews PRs, connects to remote devboxes over SSH, remembers your preferences, and wakes itself up days later to continue long-running work.
It can jailbreak your lightbulb.
And still, a power user went straight to Sam Altman on X and said the quiet part out loud: "I love Codex. But there's this one really dumb/subtle problem. Claude Code FEELS faster."
Read that sentence carefully, because the word doing the work is FEELS, in capitals, and the writer knows it. He isn't claiming Claude is faster. He's claiming Codex loses the perception war.
Then he diagnoses exactly why. "I think this is primarily because as it does each thing, it immediately slams a text update into the terminal/app, whereas Codex keeps everything contained/compressed into a line that has that little side to side fade effect."
Claude streams. Codex compresses. One shows you the work as it happens; the other hides it behind a single shimmering status line that fades side to side, telling you almost nothing about what's actually ripping underneath.
The most important part of his post is the analogy, because it's not a UX nitpick—it's a precedent. "It's almost like chain of thought—you were doing it in ChatGPT just not showing it, and when you added it people felt like it was smarter/faster."
That's the whole argument in one line. Chain-of-thought didn't make the model smarter. The reasoning was already happening. Surfacing it changed how smart the model felt, and perceived intelligence is the thing users actually respond to, pay for, and stay for.
Chain-of-thought didn't make the model smarter. Surfacing it changed how smart the model felt—and perceived intelligence is what users actually pay for and stay for.
The same trick applies here, in reverse. Codex is reportedly doing more than Claude, arguably, and showing less. So it feels like less.
His fix is modest and correct: "I think it would be worth adding a more verbose mode / experimenting with revealing more of what Codex is doing so that you viscerally feel it ripping." Viscerally. Feel. This is a request for a state layer, dressed up as a feature suggestion.
Now widen out, because this isn't one user's pet peeve. It's a pattern OpenAI itself keeps flagging, from the very top of the Codex team.
When the macOS app launched, the framing was blunt: latency, not model capability, is the primary adoption bottleneck. The specific claim is the one worth memorizing—per TBPN Digest's reporting on the macOS launch, cutting deep research from 20 minutes to 2 minutes could drive 10x more usage and willingness to pay.
Sit with that ratio. Not 10x more capability. The same capability, ten times faster, unlocks ten times the demand and ten times the revenue. Speed, not intelligence, is the near-term variable that moves the business.
Nick Baumann said a version of the same thing from the user's seat. He reportedly often runs the highest-intelligence Voice mode, but once a model has the tools it needs, speed can matter more than another small jump in reasoning quality. The tools that feel immediate win, especially with users who currently find AI too slow to bring into a live workflow.
Here's where perceived speed and actual speed collapse into the same problem. If a deep research task takes two minutes, the difference between a product that streams its progress and one that hides it behind a fading line is the difference between two minutes that feel like work happening and two minutes that feel like waiting.
Streaming doesn't cut latency. It converts dead time into visible progress. That's the exact same psychological move as cutting 20 minutes to 2—you're reducing the felt cost of the wait—except it costs almost nothing to ship.
And the stakes climb sharply once you leave the single terminal. Voice was explicitly built as an orchestration layer sitting above the Codex and Work harnesses. Each spoken task runs in its own independent thread. You can have multiple Codex threads and Work threads live at once, steering them by voice.
That is genuinely new, and it is exactly where the invisible-state problem gets dangerous. When you've forked three threads—flight, expense, calendar—and each is compressed into its own fading line, you have no felt sense of which agent is where. What's done? What's dispatched to the cloud versus running locally? What's stalled?
The confusion is already showing up. One user asked the Codex team directly whether saying "schedule this task in the cloud" from the Codex tab actually schedules it in the cloud—"It doesn't seem to work?" The replies conflicted: one said Codex can only create scheduled tasks locally, with cloud runs able to use connectors but not a local folder and local runs needing the desktop app and the machine awake, while another said the docs claim Codex or Chat can create a scheduled task.
That's not a capability gap. That's a legibility gap. The system can do the thing; the user can't see or feel where the thing is happening, so they can't tell success from silent failure.
This is the through-line. The lightbulb toolbar, the three-way voice fork, the Heartbeat that catches a fraudulent charge—all of it is real, and all of it is undermined by the same missing layer. Presence, completion, dispatch. Where is each agent, is it done, and did my command actually land.
So here's the argument, stated plainly. Surfacing agent state is not a cosmetic setting. It is the product.
The Codex versus Claude Code race will not be decided by whose model is smarter. On raw capability for general computer operation, Codex is reportedly arguably already ahead. It will be decided by whether users can feel the agents working—whether the interface converts invisible reasoning and dispatch into something visceral, the way chain-of-thought converted invisible reasoning into perceived intelligence.
Claude Code stumbled into this by simply dumping actions into the terminal as they happen. Crude, verbose, and it wins the feel. Codex chose elegance—one calm compressed line—and elegance reads as slowness, and worse, as opacity.
The frontier isn't the smartest model. It's the visible state layer for multi-thread orchestration that nobody has shipped yet. Whoever builds it owns voice-driven computer operation, because they'll be the first to let users viscerally feel an army of agents ripping through their work instead of squinting at a line fading side to side, wondering if anything is happening at all.
The future is now, exactly as Claire said. It just forgot to show its work.