Issue · 2026-09-02

Most AI news is noise.Here is the rest.

DIGEST

Fable 5.1 is out, and it's cheaper to run

Claude Fable 5.1 ships today in Claude Code and the Claude Platform, priced the same as Fable 5 with 75% cheaper API cache reads, and it gets further into a long task before needing your input (ClaudeDevs).
The direct-impact news: the release also reset the 5-hour and weekly limits for all users (ClaudeDevs).
The cost angle: Fable 5.1 solves agentic coding tasks at roughly half the cost of Fable 5 (felixrieseberg) — though an independent check says it tops the Intelligence Index but still costs ~20% more per task at max effort (ArtificialAnlys).
The real-world wrinkle: cache writes, not reads, dominate — over 65% of the cost in real Claude Code usage, so the read cut helps less than it looks (theo).
For anyone running it: /claude-api prompt-audit on your skills strips redundant rules tuned for older models (petergyang).
A concrete config recipe: set effort to low, run cost-optimize and prompt-audit, and change effort mid-conversation without a cache hit (daniel_mac8).
The Fable 5.1 system prompt leaked at 270,000+ characters — worth a skim before you write model-aware rules (elder_plinius).

DIGEST

Matt Pocockuk's bet: the smartest way to cut slop is to produce less code — force review to shrink code while keeping cyclomatic complexity low, since less code means fewer bugs to maintain (mattpocockuk).
A supporting smell: good tests rarely change, but agents keep rewriting them — they're effectively writing the same code twice (thdxr).

DIGEST

Karpathy compressed his OpenAI/Tesla years into a free 2-hour lecture running Agents → Harness → Loops → Graphs → Self-Improving Systems (elune0x).
The one-line takeaway: prompting is fading; the graph or harness layer is what survives, and "delete everything else" (AnnatarXBT).

DIGEST

A paper by Chinese researchers argues software engineering as we know it is finished, with agents replacing code as the carrier of logic (thesupermanmx).
The substance behind the title: Zhengeng Cao's thesis is that software is no longer the product (eticaprotocol).

DIGEST

steipete moved his whole team onto a shared coding harness that OpenClaw built for itself — a live case of running your own agent day to day (steipete).

DIGEST

Pi hit 100,000 GitHub stars and Pi v2 is on the way (pidotdev).
The thinking level selector is the feature to adopt: /thinking to pick a level, Ctrl+S to make it your startup default, Shift+Tab to cycle mid-session (pidotdev).