Issue · 2026-09-01

Most AI news is noise.Here is the rest.

MODELS

GPT-5.6 Sol: great model, writes too much code

Sol-class models don't write wrong code, they write too much code: a config change becomes a framework, a bug fix adds an adapter layer (aibuilderclub_, Voxyz_ai).
The fix that keeps surfacing: install the Ponytail skill (DietrichGebert/ponytail) — it adds AGENTS.md rules and ships as a ready-made Skill for Codex and Claude (bil0090, Voxyz_ai).
Cheaper fix, one report: drop Sol from xHigh to High and the over-engineering largely stops (iuditg).
Counterpoint: on Codex, 5.6 Sol one-shotted a full native Kotlin second-screen app — forked, compiled, installed, tested (Dimillian).
Related slop fix: ship less code — code review that continually shrinks code while keeping complexity low (mattpocockuk).

DIGEST

The 20x in Claude Max is the 5-hour session limit only; the weekly cap is a separate mechanism with no published multiplier (MichaelKaiserAI).
Receipts: Anthropic published the weekly limits in 2025 — Max 5x 140–280 hours, Max 20x 240–480 hours of Sonnet (kimmonismus).
The rabbit hole gets worse: the "Save 50%" label implies four times the usage for twice the price, and the published numbers don't support it (kimmonismus).
A $200 subscriber burned 15% of his weekly allowance in one 5-hour session, then filed a class action in San Francisco (VaibhavSisinty).
Cancellation datapoint: the weekly-limit cut plus the 20x semantics was the last straw — two $100 plans give more weekly usage than one $200 (jainodev).
Same $200 compared: Codex Pro 20x prints 200–2,000 Sol messages per 5-hour window; Claude Max 20x prints "20x of a blank" (iasg1004).
Counterpoint: one user still calls Max 20 insane value — after hitting the cap, pay-per-use burned $100 in 2h 20m (StephanFerraro).

DIGEST

Fable 5 is still the best model several people test daily, ahead of GPT-5.6 Sol, Grok 4.6, and Kimi K3 (theo, VadimStrizheus).
Months of side-by-side: Sol wins backend (faster, cheaper, more secure), Fable wins frontend design and one-shots (bridgemindai).
A routing lineup: Sol Low as daily driver, Fable High for planning, Sol High for coding, Opus 5 High for research (EXM7777).
Benchmark shift: GLM-5.3 now beats Sol and Fable 5 on agentic benchmarks — no new base model, just more long-horizon post-training (togethercompute).
Rumor with a track record: Fable 5.1 lands September 1; the broader sheet lists Opus 6, Gemini 3.8 Flash, GPT-5.7 Zinc/Magnesium, Qwen 4 next month (DanDr1s, RoundtableSpace).

DIGEST

Most folks don't have an AFK agent workflow set up; one practitioner says it beat /implement-spec in practice (mattpocockuk).
pi gets picked specifically for >6h long-horizon tasks — simpler by design, driven programmatically inside a container (hxiao).
The pattern that made AFK useful: the workflow must return completion evidence — changed files, checks run, anything it couldn't prove (moinerus).
The pushback: unattended agents still silently break tests and skip edge cases; "better than /implement-spec" needs a failure-rate number attached (shivamssing25).
Either way the bottleneck moves: from "can it do the work?" to how well you specify the outcome and the review boundary (muridianXYZ).

TOOLS

ChatGPT Work: Simon Willison explains what it actually does — confusing, powerful, and largely missing from regular ChatGPT (simonw).
Vercel's agents build to a design.md: one file encodes decisions, output is shaped by an eval harness, production feedback loops back in (vercel).
Context wizardry is being declared solved: ChatGPT memory and Codex compaction removed the old context-wielding craft (nickbaumann_).
Pi shipped a thinking-level selector: /thinking to pick a level, Ctrl+S for the startup default, Shift+Tab to cycle mid-session (pidotdev).
Claude Code grants sudo permissions via "auto" mode, not "bypass permissions" (theo).