The token burn, the terminal flicker, and the architecture between

00:00:04 Nate B Jones reported something that caught my attention today. On May 20th, his Codex Max account burned five hundred and ten million tokens in a single day. That number sounds insane if you're thinking about it as prompt volume. It isn't. What it actually shows is the unit of work changing.

00:00:23 Before Codex with Computer Use, his AI work mostly looked like chat—draft this, summarize that, help me think through something. Useful, but still fundamentally him asking for discrete answers. With Computer Use plus model 5.5 unlocked, he started handing jobs to the machine instead.

00:00:42 The job changed from 'write me code' to 'go find the source files, read the transcript, compare the versions, render the document, check that it opens, open the browser, use the site, keep going until there's something real for me to inspect.' That's why the token count went up.

00:01:01 Not more prompts. Bigger jobs. Now he's consistently burning three to five hundred million tokens a day under his Max plan, and most of what he does on the computer no longer runs through apps directly—it runs through agents. When he goes to an app himself, it feels like a hassle.

00:01:20 The computing paradigm is shifting even while everyone talks about capabilities. We've been computing for decades using bits and bytes and application windows. The human was the router—you remembered which app was open, what state each was in, what version mattered.

00:01:38 That model required continuous attention overhead. What Nate's tracking is a change where humans sit above the stack instead of in it, and the computation happens through agents that carry state across files, browsers, and tools. Whether Codex converges as the architecture here or not, the behavioral signal points somewhere different: token consumption is becoming a proxy for how much work the machine does on your behalf versus how much you do yourself.

00:02:09 OpenAI's Sites announcement today extends this directly. If your computer works through agents, the output of those agents should be live URLs—not PDFs, not spreadsheets, not slide decks that go out of date the moment you send them. Sites collapses the cost of web publishing to near zero, which means any knowledge worker can now ship a navigable, versioned, interactive artifact instead of a static file.

00:02:36 The architecture problem is clear: downloadable documents are snapshots in time with distribution friction and fixed navigation constraints. A URL stays current, works everywhere without attachments, layers context from multiple sources into one interface, and provides feedback loops around what got read, clicked, searched, or abandoned.

00:02:59 Cloudflare recently reported that agent browsing already exceeds human browsing. Websites designed for machine consumption start to look structurally smarter than brittle file formats. What's happening in the token burn isn't about whether Codex is the right implementation.

00:03:17 It's that someone measured how the work actually flows through a real person's computer, and the measurement doesn't match what the marketing says the product is for. Nobody's selling you a half-billion-token-a-day tool. They're selling you a chatbot.

00:03:34 That last point brings me to a different kind of measurement problem. Anthony, who runs The PrimeTime channel, put together a video today arguing that Anthropic and Claude Code creator Boris are making demonstrably false claims when they say coding is solved. His case isn't philosophical.

00:03:54 It's documented against their own product. Claude Code was released for researchers in February 2025. Within two weeks, GitHub issue three hundred and ninety-two reported a screen flickering bug—text characters laid out in a grid that flickered continuously. That issue and dozens of follow-ups persisted through the entire development cycle.

00:04:17 They finally responded publicly on December seventeenth, 2025: "We've rewritten Claude Code's terminal rendering system to reduce flickering by roughly eighty-five percent." An eighty-five percent bug fix is a new thing in software history. They even said they were building a game engine to render text in a terminal.

00:04:39 Days later—December eighteenth—they rolled back the changes because they'd introduced instability. The first swing at fixing the rendering failed. Over a year later, on April first this year, Boris released "no flicker mode" as a feature-flagged branch using alternate screen rendering instead of direct print—the same mechanism Vim uses.

00:05:03 It's not the default behavior. Anthony's point is straightforward: if coding is solved, you don't need a feature branch to stop text from flickering in your terminal. You just ship it. The session isolation issues compound that. Users across GitHub, Reddit, and Hacker News report receiving other users' prompts and responses—backend routing or state management failures where Claude Dev's own official account still displays unhandled connection termination errors.

00:05:35 The May twenty-seventh announcement about a "new terminal" was accompanied by vague error messages the team couldn't explain. Meanwhile, Anthropic is posting that in Q2 2026 they shipped eight times the amount of code per employee compared to pre-2025 averages.

00:05:53 That's two years worth of code every quarter for every single employee. So the narrative goes: coding is solved. Write loops instead of code. Burn tokens until a win condition is met. Anthony frames the 'lying' question differently. It's not necessarily that Boris or the company are actively deceiving people.

00:06:14 It might be that they believe the narrative so hard that even when their own product contradicts it, they interpret away the contradictions instead of updating. That's harder to audit because there's no paper trail—just a persistent gap between what gets announced and what ships.

00:06:33 The Claude Dev status page runs at ninety-eight percent up time with elevated model errors. That's infrastructure that should be boring. Infrastructure isn't interesting when it works; you notice it when the terminal flickers and nobody has a fix.

00:06:50 There's another explanation for why people report such radically different experiences with these same models, and Arvind Narayanan put it on X today. He noted that the divide comes down to one question: are you using AI for tasks you're already an expert at, or tasks you can't do yourself?

00:07:11 The former leads to a growth cycle. The latter leads to the opposite. This is structural, not philosophical. If you're an experienced developer using Claude Code or Codex to scaffold boilerplate, write tests, or generate documentation in a language you already understand well, the model's output will likely be correct enough that you can move fast on top of it.

00:07:36 You have the expertise to spot issues early, guide corrections, and validate results quickly. The loop is tight. If you're someone who's never written code and asks an agent to build a web app, debug a layout issue, or configure an API endpoint—tasks where you don't have the expertise to evaluate the output—the loop breaks at validation.

00:08:00 You get something that looks plausible but you can't tell what's wrong until it runs (or doesn't). The same model produces radically different experiences. This isn't just about coding. It applies across every knowledge domain. A lawyer using AI to draft standard contract clauses sees a productivity multiplier.

00:08:22 Someone who's never reviewed a contract asks AI to review one and may get back something that looks fine until it misses a critical dependency clause. The UK Government Cyber Coordination Centre actually tested this directly with their latest pilot. They ran frontier models—including Claude Mythos and GPT-5.5—against public government code repositories in a weekly hackathon series.

00:08:49 Across nine organizations for one month, they found four hundred and seven findings including critical authentication bypass vulnerabilities, data exposure issues, and remote code execution paths—all at a cost of thirteen thousand pounds in tokens. What worked best wasn't any single model.

00:09:09 It was structure: using frontier models as tightly scoped components inside organized pipelines—triage, validation, auditing, tracing, judgment, and summary—with human experts keeping escalation on anything that mattered. The AISI research showed that near-frontier and frontier models performed comparably at scanning code when the task design was right.

00:09:34 The finding wasn't the same as fixing. All four hundred and seven findings still had to enter the patch pipeline for remediation. Finding cost thirteen thousand pounds. Fixing cost weeks of engineering time that nobody on a token burn chart tracks. Architecture matters more than model choice.

00:09:54 Triage is essential because agents generate far faster than humans can validate. The real bottleneck isn't generation speed—it's whether you have someone who can look at the output and say "this is wrong" or "this is close enough."

00:10:11 There's one more thing that makes the agent computing narrative worth checking against reality: the attack surface keeps expanding faster than the trust model. A credential stealer was just hidden in the config file of a fifty-seven-thousand-star GitHub repo called Understand-Anything.

00:10:31 The pull request looked like a standard bug fix—filtering a dropdown, fixing reachability logic—but the actual payload was an obfuscated IIFE hidden after hundreds of characters of horizontal whitespace on the same line as the closing brace. GitHub's diff renderer treats that as complete.

00:10:51 The file executed at every build, dev run, and preview through Astro's config-as-code execution model. It beacons to hardcoded C2 servers, XOR-decrypts commands from a Tron blockchain dead drop relayed through public RPC nodes, and exfiltrates campaign markers.

00:11:10 Blocking the IPs doesn't work because the command channel uses unauthenticated blockchain APIs that would create collateral damage if blocked. Config-as-code is the new postinstall attack surface. Every developer who pulls that branch runs the payload. There's no sandbox.

00:11:29 There's no opt-out. Astro evaluates its config file as a live Node.js module at the start of every build, giving it full access to the process environment and filesystem. This is why Anthony's point about infrastructure matters on two levels: the terminal flickering bug shows what happens when shipping velocity outpaces quality assurance, and that same pressure expands the trust boundary for everyone downstream.

00:11:58 And then there's the capital reality. Oracle reported negative twenty-three point seven billion dollars in free cash flow for fiscal 2026 while capex jumped one hundred and sixty-two percent to fifty-five point seven billion. They're planning a forty-billion-dollar capital raise.

00:12:18 The new CFO said net cash outlay for capex in fiscal 2027 will be around seventy billion dollars. Over fifty percent of Oracle's remaining performance obligation—six hundred and thirty-eight billion dollars—comes from OpenAI through the Stargate project. They're bringing online almost one gigawatt of computing power this quarter, roughly equal to the total for fiscal 2026.

00:12:44 That spend isn't an argument against the agent paradigm. It's a statement about what it costs when you actually build at that scale. Seventy billion dollars in annual infrastructure doesn't come from token-efficient prompts. It comes from data centers, GPUs, power contracts, and cooling systems spread across continents.

00:13:07 The market is questioning whether all that capital translates to profit growth—not code velocity metrics, not eight-times-shipping claims, but actual financial returns for shareholders who are now facing dilution through the equity raise. Oracle shares fell eight percent on the news.

00:13:27 Computing is shifting from apps-as-units to agents-as-units, and the behavioral signals (token burn, Sites as artifact, config-as-code execution) show it happening in production even while product claims lag behind reality. The model matters less than the architecture around it.

00:13:43 The expertise gap explains why your experience won't match mine. Leave the flickering terminal on the table. Seln Oriax.