Issue · 2026-09-07

Most AI news is noise.Here is the rest.

MODELS

GPT-6 Astra is the new frontier default

GPT-6 Astra is the new #1 on Code Arena and reshapes the Pareto frontier as the best model at $40/Mtoken, matching latest Claude pricing (arena).
On Vercel's DeepsecBench Astra is now the best cybersecurity model, finishing in 49 minutes what Sol took ~4 hours on, at nearly the same cost and 50% cheaper than Claude Opus 5 max (rauchg).
ARC-AGI-3 jumped from 66% on the standard harness to near 100% with a continuous-conversation harness and custom compaction — a step change for interactive reasoning (fchollet).
In 100 complex multi-agent coding environments Astra won by a landslide, even more dominant than the Fable 5 release (leo_linsky).
The release cadence is the story: only 56 days between GPT-5.6 Sol and GPT-6 Astra, versus 877 for GPT-4 to GPT-5 (lochan_twt).
It rolled out to Pro, Enterprise, and Business Premium in Work/Codex plus the API, with Plus and Business users next (sama).
List price is $10/$50 per M in/out, matching Fable 5.1 — but the screenshot-worthy 99.9% ARC-AGI-3 number needs its footnote read (bykimdohoon).
Not everyone is sold: after 24 hours one report found Astra overhyped for branding, copy, positioning, and conversion work (scheemunai).
Another tester calls it a show horse, not a workhorse — great for demos, not yet the daily engineering model (kieranklaassen).

CONTEXT

Astra's compaction is experimental and off by default; a short prompt enables it so notes are saved across context windows and earlier messages and tool calls are searchable (daniel_mac8).
In Codex, Astra can keep notes across windows and retrieve accumulated details instead of repeatedly compressing them when the window fills (gabrielchua).
The change is in config.toml: compaction used to summarise and drop why a fix failed; Astra now keeps notes and can search earlier tool output (JustinMiddler).
Cost caveat: the 1.05M-token window is real, but prompts above 272K input tokens are billed at 2x input/cache and 1.5x output (DonizetiFerre16).

ORCHESTRATION

A Spotify engineer's Claude Code plugin blocks big file reads at the hook layer and reroutes them to a cheap Gemini worker — routing advice in CLAUDE.md alone was ignored (builderleader).
Claude Code shipped /skill-doctor, which shows which skills load, which go unused, and how many context tokens each costs (dani_avila7).
For AI-authored PRs, @dexhorthy's /show-me renders readable diffs; combine with pseudocode descriptions, Mermaid diagrams, and evidence of TDD (mattpocockuk).
Pi extensions remain a model to copy: one git-commits a checkpoint before every turn, adds a run_tests tool, and an /undo command (pidotdev).
Grok Bot is directionally ahead of current harnesses: topic areas with a main orchestrating agent instead of threads is a pattern worth stealing (RhysSullivan).

DIGEST

When Astra burns Plus quota too fast, let Sol do everyday coding and testing and bring Astra in as an advisor for the hard parts (Voxyz_ai).
Astra was reportedly trained end-to-end for multi-agent orchestration; the astra-advisor plugin gives it a framework to delegate to any OpenAI model (daniel_mac8).
Audit skills and AGENTS.md before an Astra session — it does much better without GPT-5.x-era context bloat (daniel_mac8).
The advisor only helps if it doesn't get the whole transcript: pass the diff and a failing test so it reads code like a fresh reviewer (DevCalledFede).
A concrete setup: Sol as builder and Astra as advisor in Codex, two panes that communicate with each other (kevinkern).