Kelley Ignites the Bun Drama, Codex Becomes ChatGPT & GPT-5.6's Price Shock
The Zig Post Fallout
Yesterday Jarred Sumner shipped the "Rewriting Bun in Rust" post; today Zig creator Andrew Kelley published his response — "My thoughts on Bun's Rust rewrite" — and it became the day's angriest thread-storm.
Theo read it live on stream and detonated (post, 83K views): "I've never been so pissed off on camera. Andrew is a petty, evil man, and you should not trust any software he writes." Pushed on whether that's a stretch, he doubled down: "If he goes after his users in such a vile way, I don't think you should use his software. He is a ticking time bomb." The replies carry useful context: the resentment reads as being about Jarred taking the VC path, Zig banning AI contributions (Bun was Zig's flagship project and its biggest donor — reportedly $60k/yr, now ended), and the post repeatedly questioning Jarred's engineering judgment. Plenty of pushback too, from "petty sure, evil?? why?" to a pointed callback to Theo's own ffmpeg-maintainer bounty episode.
Charlie Marsh's cooler-headed version (post, RT'd by theo): "I find this post to be really disappointing... It's explicitly framed as 'not a personal criticism', then goes on to include a bunch of very personal criticisms about Jarred."
Armin Ronacher, notably not a Bun fan, defended Jarred (post): "what @jarredsumner is doing is pretty fucking awesome and I can confirm that he's great to hang out with... The responses to his blog post by some people are uncalled for." Best replies: Jamie Turner's "It's reprehensible and condescending. Call me crazy, but if I lose my biggest customer, I'm gonna consider it my fault", and Seb Insua calling it "one of the most poisonous posts I've ever seen... It's sad that some people's reaction to AI makes them treat other humans like this" — because underneath the personal feud, this is really a proxy war over AI-assisted rewrites (Zig forbids AI contributions; Jarred used Claude to port the entire runtime).
The contrarian minority is worth reading too: Ben Vinegar's "we've gone soft" (72 likes) and a reply collecting the Zig side's actual grievances — correctness advice ignored, and AI-generated changes that allegedly made the Zig compiler non-deterministic.
Codex Becomes ChatGPT
The other half of Wednesday's OpenAI launch — the Codex desktop app being folded into a unified "ChatGPT" app alongside the new ChatGPT Work — aged badly overnight.
Theo: "turning Codex into ChatGPT Desktop is a generational fumble" (post, 304K views, 4.4K likes, 483 replies). The replies are a catalog of migration pain: the Codex app vanishing from people's Macs, accounts not rolled over so threads and workspaces are unreachable, Shift+Tab planning mode gone, and kitze summarizing the brand damage as "aura loss of biblical proportions". Recurring take: "It is Cowork/Code all over again" — several people note OpenAI copied the one Anthropic decision devs disliked. OpenAI's dkundel offered the defense: "We just made this setting a dropdown. It's the same Codex".
Simon Willison speaks for everyone (post): "I am so confused by ChatGPT v. ChatGPT Codex v. ChatGPT Work v. Claude v. Claude Code v. Claude Cowork right now!" He read OpenAI's own help page on the split and reported back: "This did not help".
The contrarian case, from Jerry Liu (post): he actually tested both and prefers the ChatGPT Work/Codex split to Claude Cowork/Code — the interface is more unified, chat history is shared between modes, and "the team put more care into the specific toggles that should feel different." He'd still rather OpenAI went all the way to one app surface for all work.
Theo's bigger lament (post): "Codex just rebranded to ChatGPT. Claude is spamming me with Excel plugins. xAI is bragging about how good Grok is at PowerPoint. I knew the 'AI labs focus 100% of their effort on software devs' thing wouldn't last, but it's sad having it end all at once." Corey Quinn adds the money angle (post): nobody can pin down when GPT-5.6's price jumps 50% or when it stops being included in subscriptions — docs, blog posts, and employees all disagree.
Practical scrap for the migration: keep a "Codex" alias so Spotlight/Raycast searches still work.
GPT-5.6 Day Two — Benchmarks & Economics
The launch-day dust settled into numbers, and the story everyone converged on is price-performance, not raw capability.
Simon Willison's full write-up is out: Notes on GPT-5.6 (post) — the API additions he flags as most interesting are programmatic tool calling and first-class multi-agent support, plus 18 pelican SVGs covering all 6 reasoning levels × 3 models (Sol, Terra, Luna). A sharp reply worth keeping: programmatic tool calling means routing decisions leave the context window — "deterministic execution, no token burn on deliberation, testable behavior. The three things agents usually sacrifice". OpenAI's own livestream featured a 3D pelican riding a tricycle, bicycle, pony, and another pelican as a nod to Simon's benchmark. One field report on Terra to balance the hype: it kept stopping mid-task until the user rigged a timer thread reminding it every 30 seconds to keep working.
The benchmark sweep: OpenAI says Sol sets a new SOTA on the Artificial Analysis Coding Agent Index at 80.0 — 2.8 points above Claude Fable 5 while using less than half the output tokens, in less than half the time, at about a third less cost (announcement). ARC Prize reports Sol is the first verified frontier model to beat an ARC-AGI-3 game (7.8% SOTA, "best model at orienting in a situation it's never encountered"). And on CursorBench the framing that stuck was "the economics are brutal for Anthropic": Fable 5 Max 70.5% at $17.32/task vs. Sol Max 67.2% at $5.22 — 95% of the performance at a third of the cost. LLMJunky's take (post): "unless you have a task that only Fable can do well," the price story decides it.
Not every benchmark moved: LlamaIndex ran GPT-5.6 through ParseBench day-zero and found no change vs. GPT-5.5 on document understanding — still strong on text and tables, still struggling with charts and layout. And LLMJunky's Rocket League one-prompt demo verdict: decent, "but Fable mogged here".
The "work at the level of your ideas" demos: Sharif Shameem says his favorite thing about 5.6 is that it's "a fucking stellar researcher" and posted the entire prompt used to get Sol to post-train Luna (post, 172K views) — a ~500-word prompt, though so heavily redacted the replies turned it into a meme ("more redactions here than the Epstein files"). In the same genre, skirano vibe-coded a language model from a single prompt: 5.6 built the training pipeline and trained a model from scratch on his iMessage history, locally on a Mac.
Theo's month with Sol: he burned over $200k in tokens with gpt-5.6-sol over the past month and released a video overview of everything he built with it (including a Rust port of his Hermes agent), with a "proper review coming soon" (post + video). The replies split between awe and "$200k in tokens and still no review, that's a sponsorship not an opinion" — the who-paid-for-the-tokens question goes unanswered. Also his succinct standalone verdict: "It's a good model".
Claude Code & Anthropic Updates
Anthropic reset everyone's rate limits: Thariq's "Enjoy more Fable!" (398K views) quote-tweets the announcement that 5-hour and weekly limits were reset for all users — read universally as a competitive response to the GPT-5.6 launch ("Fantastic competitive energy you all"). The replies also pin down the elephant: people are still treating July 12 as the date Fable access via plans expires, and Thariq didn't correct them. Relatedly, kitze notes Anthropic is including Fable in the Max plan, and Theo marks the date: "Fable 5 released 1 month ago today".
/checkupkeeps climbing: Boris Cherny's announcement from Tuesday is now at 720K views / 11K likes (covered in yesterday's roundup). New in the replies since: confusion about/checkupvs./doctorpersists, and the sharpest critique is that auto-cleanup makes your setup Claude-Code-only — "since Anthropic refuses to support standards for skills or AGENTS.md, I won't be using this".Armin Ronacher's Anthropic "announcement" (post): some aggregator listed him as an Anthropic employee, so — "I guess I got to announce that I'm leaving Anthropic." (He never worked there.)
Agentic Coding & Craft
pi 0.80.4/0.80.5 ships cache-miss accounting (Armin Ronacher, changelog):
/sessionnow shows how much money you lost on cache misses, with cache warnings in the transcript — plus GPT-5.6 support, new extension and lifecycle hooks, and a bigpi configupdate. The design insight in the replies: "Seeing the wasted-money line inside /session is what changes behavior. People fix what stares at them in the transcript". One user reports pi spending about half the tokens of Codex for the same output on the same model. Armin also says pi is now investing in grammar-constrained decoding and dynamic tool registration, arguing those areas have finally stabilized enough across providers to build stable APIs on.Matt Pocock recorded the missing tutorial (post + video, 240K views): his skills repo hit 160K stars and 7.5M downloads without one, so this walks the whole flow end-to-end —
/grill-with-docs→/to-spec→/to-tickets→/implement→/code-review(all skills here). The new/wayfinderskill (deliberately excluded as "not entry-level"; tutorial in a couple of weeks) is generating its own buzz: one user describes it as a genuine break from the Research→Plan→Implement paradigm — "your agent can now be both a PM and a developer" — though another spent a whole day watching it spawn dozens of GitHub issues without asking a single question, which Matt is treating as a bug report.Matt's personal-wiki experiment is compounding (post): every weekday he now gets a 15-minute generated podcast covering what flowed into his wiki from X, Discord, Slack, Gmail, and GitHub issues — doubling as a vibe check on the pipeline itself. Related aphorisms from his feed worth keeping: "An agent is just a model, harnessed, in an environment", and AI-written PRs/commits being easy to review is an alignment problem, not a model problem — shared vocabulary from domain modeling makes commit messages dramatically more readable.
Thariq's one-liner (post): "one of the core skills of agentic coding is reducing your unknowns."
Quick Hits
swyx on AI-era SEO (post): "whoever does AEO for @resend needs to get a raise, all the leading frontier models keep trying to use resend for emails even when I have existing transactional infrastructure already set up." Answer-engine optimization — getting your product to be the one models reach for — is quietly becoming a growth channel.
Codex housekeeping trick (LLMJunky): ask Codex to clean up your stale threads — traditional automations reuse the same title every run, so they're easy for the model to identify and sweep.
GPT-Live API design partners wanted: Greg Brockman is soliciting design partners for OpenAI's GPT-Live API — creative new apps or integrations into existing products.
Cognition stakes out its lane (Jared Zoneraich, RT'd by LLMJunky): "Cognition only cares about software engineering. We're not trying to build an AI God... just a really great AI Software Engineer" — a pointed contrast on the same day OpenAI pivoted its dev tool toward general work.
Compression claim to watch (post): a purported breakthrough running models "losslessly" on 13% of the usual hardware, with calls for someone to validate it with a full terminal-bench-2.1 trace. Extraordinary claims, no trace yet.
The Theo–Charlie Grahn saga continues (post): Theo reposted the year-old video that's still drawing takedown attempts, then reported another copyright strike with identical content and timestamps — "Despite being informed that these are invalid and malicious. Willful ignorance. Probably illegal."