X Post Facto
Fable 5.2 lands faster and cheaper beside a prompt post, a Qwen4 preview, and agents that don't learn
X Post Facto · 8–14 Nov 2026
X Post Facto is our weekly pause on the feed: a look back at the AI posts that stood out, what they said, and why they’re worth a closer read.
This selection is a short, practical week. Fable 5.2 arrived faster and cheaper in the Claude Platform and Claude Code. The same day, Pliny the Liberator posted system instructions and tools for that model in Claude Code. Qwen said Qwen4 is coming and put Qwen4-Max in preview. The other notes are about living with agents: cloud simulators built for them, a new benchmark where three named agents fail to learn, and a line from Theo that T3 Code should not burn sub limits faster than Claude Code.
A price cut, then a very long prompt
ClaudeDevs said Fable 5.2 is now available in the Claude Platform and Claude Code. It is up to 30% faster than Fable 5.1, at $8/$40 per 1M tokens, down from $10/$50. That is release news you can put next to a budget: less wait, lower prices, on the two surfaces the post actually names.
Later that day, Pliny the Liberator marked it a system prompt leak and shared "the full system instructions + tools for Fable 5.2 in Claude Code," with all 950,000 characters on his GitHub under CL4R1T4S. A Claude Web App prompt, he added, was "comin up next."
Availability, twice, and not the same kind. One post makes the model cheaper to run. The other puts a very long steering text in public, under a name — CL4R1T4S — that is doing its own kind of advertising. Neither post ties the price cut to the leak. If you use Claude Code, they are still the two things worth reading before you treat 5.2 as a quiet drop-in.
Qwen4, mostly still arriving
Qwen said Qwen4 is coming: "Deeper knowledge. Richer world understanding. Stronger coding and agentic capabilities." The sentence that names actual work runs "from complex coding and knowledge work to front-end builds and professional software workflows." Qwen4-Max is now in preview on Alibaba's Token.
Coming, plus a preview. Fable's post came with a speed figure and a price. This one comes with a destination and a Max the account says is in preview. The adjectives are the trailer. The preview is what the post says is available now.
Bring the recording back first
Expo opened on the bottleneck: "Coding agents write mobile code faster than any team can review it." Expo Simulators are "cloud simulators built for agents." The loop is an order of operations. The agent runs its change on a cloud simulator or emulator and "brings back the recording before it asks for your attention."
That last clause is the product. A recording is not a review, but it is a better interrupt than a diff you have to imagine on a phone. If agents are outrunning mobile review, the argument worth having is not whether another simulator exists. It is what the agent must carry back before it taps a person.
Three names, nothing kept
Daniel Lenton was not celebrating a ship. OpenClaw, Hermes Agent, and Prime Agent, he wrote, do not meaningfully learn from experience. On Continual-ARC, "our new benchmark at @letsunifyai," none of them were able to memorize anything of value and none were able to generalize to new tasks. Training a small 9B model's weights did both.
The split is the part worth sitting with, and it is narrower than a slogan about agents. He named three systems and a benchmark he is introducing. On that test, they neither memorized nor generalized; weight training on a 9B model did both. Those agents, in this result, did not stand in for changing the weights.
If it costs more, he wants the details
Theo drew a bright line: "T3 Code should NOT go through your sub limits faster than using Claude Code directly." If it does, he called that a bug. A ton of small issues have been squashed, he said, T3 Code should be at parity, and he asked people to report and share details if they see otherwise.
The useful standard is hiding in the caps lock. A path through T3 Code should not be a more expensive way to spend limits than Claude Code used directly. If it still is, he does not want a vibe. He wants the details.
If you run Claude Code, read the Fable 5.2 note and Pliny's prompt post before you assume the only change is speed and price. If the question is whether agents keep anything, start with Lenton's Continual-ARC post. The rest of the set is linked below.