Did Codex Overtake Claude Code? Ultra Mode Demystified & the Tokenmaxxing Backlash

Did Codex Overtake Claude Code?

The pricing drama has curdled into a scoreboard question. Tibo Sottiaux thanked "the 7M active users now using Codex and ChatGPT Work" and dropped a banked reset on every account to celebrate — and Latent Space did the arithmetic in public: "Did… Codex just overtake Claude Code? 24.5 hours ago Tibo announced 6M active users. This means Codex usage jumped 1M in ~ONE DAY. The last user number we heard from Claude Code was 2M in Feb. Very big if true." Asked what caused the jump, swyx's read was "pretty much delayed word of mouth + cost-vs-performance results coming in."

The replies did the sober work of poking holes in the comparison — and they're worth reading as a corrective to the hype:

Ultra Mode, Demystified

The single most useful post-mortem of the weekend usage carnage came from Theo, with "'Ultra' mode is not what you think" (53K views, 596 likes) — a deep-dive video on what Codex's newest mode actually does. The mechanism, now widely understood thanks to the thread: Ultra fans out subagents that each carry the entire context window, so running everything on Ultra silently multiplies token spend. The confessions rolled in — "WHAAAAAAAAAAT? I burned 80% of my 20x tokens yesterday because I ran everything in Ultra, in one day!" — as did the retraction from a former skeptic: "As much as I hate Theo, you guys shoulda listened to him day one. I felt like 5.6 Sol on Ultra was bad and overhyped and missed the results of 5.5 xhigh. Turns out it was going in circles and over-complicating things because Ultra." The sharpest design note: "I like v2 in theory — subagents talking to each other instead of just the root agent could be really good. But they don't need the ENTIRE context to do so." (And the perennial Theo review of the video itself: "bro, your videos are in --verbose-ultra mode. Can we get a --short version?")

A smaller but practical Theo TIL worth stealing: "most people didn't know Codex has a question tool — because they gate it to plan mode for some reason??" His broader mood on the churn, though, is upbeat: "the web dev world is adjusting to AI fastest because we're used to everything changing every few weeks."

The Tokenmaxxing Backlash

Underneath the leaderboard energy, a genuinely human moment: a user, Rob Hallam, posted that he'd ended up in hospital from stress after staying up all night pushing his Fable limits, thinking access would be pulled — "Health comes first. Do better @AnthropicAI." Armin Ronacher's response became the line of the day (29K views, 651 likes): "One thing I know for sure: if you end up in hospital from stress because of tokenmaxxing and FOMO, you're ngmi." The replies mostly agreed the culture has tipped into something unhealthy — "if I were feeling so bad I'd go to the hospital, my confidence that the work I did wasn't a slop-fest is 0%" and "Fable is slow — I have no clue why you wouldn't just go to sleep while it's working. What was the guy even doing, monitoring its thinking tokens?" — with one calling it "the most accurate usage of 'ngmi' I've ever seen."

The reality check under all the "unlimited" bravado came from dax's poll (RT'd by Ronacher), "None of you have unlimited tokens": of 6,635 votes, only 22% had truly unlimited access via work, 47% limited, 31% none. Ronacher's companion note: "I opted GPT 5.6 out of my personality steering. It's brief enough out of the box." — and, on the laziness spectrum, "How lazy can you be? This lazy. I added a keybind to send 'continue.'"

Delete the Docs

Matt Pocock's provocation was the week's biggest craft thread (273K views, 2,356 likes): "1. Delete the docs you create to explain your code. 2. Take the tokens you save on updating those docs. 3. Spend them on making your code self-explanatory." The best reply cut through the novelty: "Replace the word 'tokens' with 'time' and it's 2020 again" — to which Pocock happily agreed: "Yep, nothing changes." His carve-outs matter, though — the two things code genuinely can't explain about itself: alternatives considered and domain language, i.e. ADRs and glossaries, plus navigation pointers — "thin chains of links that form the 'roads' the AI uses to navigate your codebase." On keeping ADR history from rotting: "mark old ones as deprecated, and maybe move them out of the repo so they can't be searched for."

The counter-evidence came from Sentry's David Cramer (@zeeg), whose result Pocock retweeted rather than dodged — "I've kept specs colocated in the repo for the last N months. I don't think it's adding value. I spend a lot of compute keeping them updated (they basically become docs), and they didn't help enough with steering. Many times I caught them outdated-enough that the agent misunderstood what truth was." The convergent conclusion across both: rely on the code, keep only the thin layer of docs the code can't produce itself.

Separately, Pocock keeps generalizing his /grill-me skill so it works "anywhere, not just engineering" (design tree → decision tree, codebase → environment), after someone reported using it in technical interviews to watch whether the AI drives the candidate or the candidate drives the AI.

Agents Leave the Editor

The throughline across several feeds: agents are outgrowing the code editor. Lee Robinson framed it cleanly — "Coding with AI showed you could give models powerful tools and dramatically improve usefulness. We're seeing the same thing now for all knowledge work: the agent can use the computer as you would and gain context from all the apps you use." He handed out two pieces of credit: "OpenClaw for making file-based memories widely used, and Codex for popularizing local computer use in the past few months."

steipete's cloud-agent experiments got genuinely strange — in a good way. After moving work to the cloud: "I moved our maintainer agent to the cloud and they are fighting already." And the emergent-coordination anecdote worth quoting in full: "GitHub was flaky so Codex had issues merging many stacked PRs/sessions and started talking and coordinating what gets merged when, autonomously. I didn't ask it to, but one session eventually took over coordinating all the others and solved the issue." This lands the same day OpenClaw shipped v2026.7.1 — 3,063 contributions from 532 contributors, with Web UI/onboarding overhauls and major iOS/Android/macOS app work. His one-liner prompt tip: "'stress test' is a good prompt."

The most concrete "which model when" artifact came from swyx's Big-Boy-project assembly line: "sol ultra to plan → fable 5 to critique → sonnet 5 / terra ultra / swe 1.7 to ultracode → devin review to review (using kakuna). Always use a variant of @mattpocockuk's grill-me or @trq212's interview-me to elicit decisions upfront." He also zoomed out on the coming glut: "By end of year we should have GPT 6, Fable 5.5, Gemini 3.5 Pro, Grok 5, Kimi 3, GLM 6, DeepSeek v4.5… Never has the frontier been so multipolar. The benefits to agent orchestration / LLM-council judges are ramping up. Invest accordingly."

Jerry Liu's take rhymed with swyx's: "Massive opportunity for startups to build specialized, domain-specific workflows at the Pareto frontier of accuracy, cost, and latency. Not every task requires frontier intelligence (e.g. 5.6 Sol xhigh) — every task has an optimal model+harness config." His verdict on the week: "OpenAI is absolutely cooking this week" — quote-tweeting sama's pointed jab at Anthropic, "come for the best model, stay because we don't treat you with contempt."

Videos

  • Theo — "'Ultra' mode is not what you think" — the deep dive that finally explained why Ultra was silently nuking weekly quotas (subagents inheriting the full context window) — via @theo.
  • State of Agentic Coding — Armin Ronacher and Ben Vinegar's latest, this time with special guest Mario Zechner (@badlogicgames): vibe-checking Fable & GLM 5.2, an RL primer from Armin, and how harnesses smooth over model jank — via @mitsuhiko.
  • Matt Pocock's /wayfinder live demo — a 75-minute stream ("WAY too big for a 75-minute stream, holy crap") walking through the codebase-navigation skill — via @mattpocockuk.
  • "From fork() to Fleet: Designing an Agent Sandbox Cloud" — swyx's favorite AIE sandboxing talk, from the track that went from a "trial balloon" a year ago to a packed room — via @swyx.

Quick Hits

Note: @potetotes' feed again returned no items (Nitter serves an empty channel for the account), so it's unrepresented in this dispatch. @karpathy and @bcherny had no new posts in the window.