Sol Defects to Claude Code, Fable's Deadline Eve & Claude Gets a Browser
The Harness Flip — Sol in Claude Code
Theo spent the evening wiring gpt-5.6-sol into Claude Code, live-tweeting the experiment from "Trying something real quick (look closely at the top)" through "Holy shit it's working" to the conclusion that broke containment:
"gpt-5.6-sol is meaningfully better in Claude Code than in Codex. I'm going to crash out so badly over this" (post, 138K views, 2.2K likes). The specifics: it makes better designs ("WHY. WHY IS IT BETTER AT DESIGN WHEN USED IN CLAUDE CODE"), Claude Code's workflows use way fewer tokens than Codex Ultra for the same output quality, and Claude Code's subagent orchestration is "significantly better than Codex's… Turns out code is good, and sol is good at writing code!" Asked whether this changes his opinion on custom harnesses: "Kind of? I have very complex thoughts here. Video coming soon. Codex fork hopefully not coming after but it might have to..."
The replies are a snapshot of the harness-hopping era: "It's kinda ironic at this point that for almost every lab's models, they tend to perform better in other labs' harnesses", "Harnesses get WAY less attention from labs than they should", and the counter-report that Fable is meaningfully better in Codex for backend/long-running work — the flip apparently goes both ways.
The concrete Codex fumble that started it: if you set gpt-5.6-sol to "ultra", every subagent it spawns is also Ultra (186K views) — the
spawn_agenttool doesn't let the model choose effort level, so quota evaporates three layers deep. Best war story in the replies: a user who ran Sol Ultra overnight and woke up to an $800 auto-pay surprise and a folder full of spec files — "the red-team subagent spent 8 hours finding increasingly ridiculous non-issues in the scaffold. Zero actual code built." Theo's one-word reply: "Accurate." Rambone's summary deserves framing: "Subagents should default to Luna. Why are we paying the PhD researcher to do a grep."There is a partial workaround — hidden config flags. Eidzoku shared the
[features.multi_agent_v2]incantation (hide_spawn_agent_metadata = false,tool_namespace = "agents"), which Theo confirmed is required before subagent metadata even passes through, while dismissing the "just define roles in config" defense: "The whole point is that the model can right-size the subagent orchestration depending on the task." Codex's v2 subagent implementation is still cooking, and 5.6 seems to trigger an auto-upgrade to it — "a rat's nest." OpenAI's eric provencher replied "We're working on this"; Theo: "Please do. I want workflows so badly."
Fable's Deadline Eve
Tomorrow is July 12 — the announced end of Fable 5 access in Claude Code plans — and the mood shifted from mourning to game theory.
Theo's take got 4.7K likes: "I would like to thank OpenAI for putting out a model exactly good enough to force Anthropic to keep bundling Fable in the Claude Code plan" — with the caveat that the cutoff is still probably happening, "but it would be insane for them to go through with it."
The replies are a betting market: Polymarket gives an extension 37% (Michael Guo thinks it's higher), one reply calls 95%+ odds of an extension by the 13th "from a game theory perspective," another predicts they'll keep adding a week at a time until the next Fable model is ready. The maximalists want more: "gonna need them to allow fable for 100% of the usage limits now" (currently capped at 50%), and "If they go through with it prepare for mass dumping of Claude accounts". The defection case is also represented: "OpenAI has been very generous with their usage resets… It has also caught and fixed several things that Fable 5 did not. I will not be paying extra for Fable 5."
Claude Code Gets a Browser
Claude Code on desktop now has a sandboxed in-app browser (ClaudeDevs announcement, 1.7M views, 16K likes): Claude can pull up docs, designs, or any site, read and click through them the same way it interacts with your local dev servers, and you choose whether sessions persist. RT'd by both Boris Cherny and Thariq. The replies capture the range: "watching it click through docs and actually USE them is wild", the timing jab — "takes special kinda balls to release this the day after OpenAI sunset Atlas" — and the counterpoint that OpenAI's new ChatGPT app shipped its browser with password/cookie import a day earlier. One practical gotcha already surfaced: it apparently broke someone's Playwright MCP setup in the desktop app.
Simon Willison reads the Atlas retirement as a category ending (post): "With Atlas being retired in favor of the browser embedded in the ChatGPT app I wonder if the whole category of AI-enhanced browsers is coming to a close. The security/privacy issues remain unsolvable IMO — I want my AI to use its own separate browser and stay out of the one I use." Notably, that's exactly the architecture both labs converged on this week: the agent gets its own sandboxed browser rather than driving yours.
Codex vs. Work — The Confusion Continues
Simon Willison kept pulling on yesterday's ChatGPT-app thread and got somewhere useful:
Is Codex a strict superset of Work Mode? (post, 81K views) The crowd consensus: yes for engineers — "work mode is for anyone who is intimidated by git" — and OpenAI insiders say the split is purely UI, not capabilities. The most substantive defense of Work Mode's existence: enterprise employees often literally can't install git, and a VM-sandboxed "Work" surface is what gets ChatGPT deployed org-wide instead of engineers-only.
Except it's not purely UI: Simon found that in the mobile app, Work mode can run code that talks to the Internet while Chat mode can't — he had both modes try
yt-dlpto extract YouTube subtitles; it failed in Chat and worked in Work. Also noted: Sol on Medium might be the new default for coding (vs. his previous favorite, 5.5 xhigh), and the app is bringing chats and projects back into the sidebar — "hopefully that means classic chat won't hide in that weird little floating window any more."
Agentic Coding & Craft
pi lands dynamic tool loading without cache wiping (Armin Ronacher, 431 likes): the next release loads tools mid-session without invalidating the prompt cache, with "somewhat consistent API behavior between OpenAI and Anthropic." The mechanics, from his replies: adding tools works, removing them still wipes caches ("just fundamental limitations"), and the use cases are progressive disclosure of tools plus clean state transitions (Plan → Implement). He tried a generic
dynamic_callfallback tool but found smaller models handle it poorly, so DeepSeek falls back to cache-trashing. Best reply-summary of why this matters: "Give the model 200 tools and it guesses. Give it the 12 it needs right now and it behaves more like a coworker." Related Armin envy: "I would kill to know what the codex backend is doing."Matt Pocock's durable case for skills (post): "To me it feels like I'm trying to build good software practices, and the distribution mechanism is skills. My bet is that good software practices haven't changed and likely won't change very much." The sharpest reply cuts the other way: one team had to redesign most of their skills after Fable shipped because they were "actually making the model dumber" — coding skills may get absorbed by the models ("3 months ago telling a model to tdd was relevant, nowadays not so much"). Counter-signal from another reply: 5.6 Sol proactively invokes skills better than any other model, which makes good skills more valuable. Matt also shipped a run-once skill that sets up deep-module package boundaries in any TypeScript codebase (imports only via
<name>/index.ts), is considering his own skills-install CLI for per-agent customization and dependency tracking, and is polling on an official slide-deck template for people giving internal talks about his skills.What 2,300 tokens/sec feels like (LLMJunky): he built Damage Scout with the Gemma team — Gemma 4 on Cerebras sampling frames from a rental-car walkaround video and rendering an annotated damage report in under 6 seconds. "You have to be much more deliberate about your prompting because there's no steering 2,000 TPS… Tool calls become the bottleneck." His tip for trying fast models in a coding agent: use OpenCode, since Cerebras is stateless and OpenCode manages session history for you.
PSA of the day: after someone's agent trashed files outside the repo, dabit3's widely-RT'd fix is a global PreToolUse hook that hard-blocks any command touching paths outside the repo (works even in yolo mode), with
permissions.denyrules as backup. LLMJunky demonstrates the no-docs version: "Hey Codex, create a global PreToolUse hook that blocks destructive commands" — "Do not RTFM. Just PTFC."
The Zig Fallout, Day Three
Armin Ronacher on the tribal turn (post, 436 likes): "Shitty personal attacks in a blog post are one thing, but when people reply 'now I like zig even more' really makes me wonder what went wrong in our communities." His summary of the affair: "Zig creator is unhappy that Bun moved off Zig and decided that the best response… is legitimate disagreements mixed with airing grievances and personal attacks." Several replies read the anger as displaced AI anxiety — "People lash out esp. at people linked with Anthropic et al. Uncalled for in this instance but that simmering anger is there" — and the pithiest diagnosis: "Programming languages are politics for computer nerds."
antirez weighs in (post, RT'd by Armin): "People for years told me why I never talked about competitors of Redis. The Zig / Bun drama is a good warning. Do your job, without thinking everything is some form of competition or comparison." And for comic relief, Aaron Stannard's outsider's guide to the drama (RT'd by steipete): "Bun == thing that crashes OpenCode & Claude Code. Rust == artisanal programming language for furries. Zig == even more artisanal programming language…"
Quick Hits
Sol/Luna dominate the cost curve: Artificial Analysis charts show GPT-5.6 Sol and Luna ahead of Terra at every point on the Intelligence vs. Cost-per-Task frontier, with Luna the standout on cost efficiency.
Prompting 5.6 Sol is different: eric provencher's guide, "a lot of people are still prompting the model exactly as they did 5.5… 5.6 Sol is a lot more tenacious and thorough" — with an official prompting guide on learn.chatgpt.com.
Codex app "More Details" (LLMJunky, RT'd by steipete): highlight any text in the Codex app and a popup uses GPT-5.5 Instant to explain the highlighted reasoning/concept in more depth.
Sol as a CAD partner (adamdotnew, RT'd by LLMJunky): GPT-5.6 Sol building a 25-DOF dexterous robot hand in Fusion — 12 tendon spools, 16 QDD actuators, fully parametric — in 69K tokens and 7 minutes. "It's now our default model."
Should AI engineers still read code? (video via altryne, RT'd by swyx): his AI.Engineer talk on the "ZL Continuum" — the divisive read-your-code debate that pulled in Theo, Mitchell Hashimoto, Primeagen, and Uncle Bob. Also from the WF26 track: Geoffrey Litt's Design Eng keynote ("I think it's important for people to understand how code works"), and Matthew Berman's sit-down with swyx on AGI, frontier-lab strategy, chips, and Fable.
OpenAI doubles its bio bug bounty (announcement): the Bio Bug Bounty becomes an ongoing private program with rewards doubled to $50K, targeting researchers probing biology safeguards.
Theo's software-churn warning (post): "There is likely going to be mass revolt around software devs due to the amount of software churn in the AI era. We can stop this. We probably won't though."