Issue · 2026-09-04

Most AI news is noise.Here is the rest.

MODELS

GPT-6 Astra changes the model-routing calculation

OpenAI launched GPT-6 Astra for computer use, coding, science, and security, but the rollout remains staged (sama).
ARC-AGI-3 rose from 66% with the standard harness to nearly 100% with continuous conversation and custom compaction, showing that harness design still matters (fchollet).
Astra matches Fable 5 on the Coding Agent Index at under half the cost, but costs 2.5 times GPT-5.6 Sol per token and 75% more per general task (ArtificialAnlys).
Early use puts computer and browser control first, Codex as the daily driver, and Fable 5.1 ahead for large architecture work (alliekmiller).
A detailed test reports stronger migrations, SQL reduction, FFmpeg optimization, computer use, subagent control, and useful low-effort output; token efficiency versus Fable remains open (petergostev).
A practical first task is to let Astra find repeated prompts, manual steps, missing skills, stale project instructions, and stalled workflows, then give execution to cheaper models (shannholmberg).
Databricks reports a new high on OfficeQA Pro and stronger enterprise data reasoning and document understanding, with Astra routing planned for its gateway (ivanzhouyq).
AutomationBench scored Astra at 41.4% versus 28.8% for GPT-5.6 Sol; importantly, Astra stopped instead of guessing when it could not find the governing instructions (wadefoster).
One tester confirms Astra can reverse-engineer binaries without source code and use the result to optimize or repair applications (mweinbach).

ORCHESTRATION

Gemini 3.8 Flash is now available in Antigravity for long-horizon coding and autonomous-agent work (antigravity).
Google positions it as a stronger workhorse for complex agentic and multi-step tasks (GoogleAI).
Conflicting Terminal-Bench 2.1 and 4.0 placements are a warning against routing work from one benchmark version (LexnLin).
Vercel AI Gateway now serves google/gemini-3.8-flash at 50% off through December 31 (vercel_dev).
Artificial Analysis measured better coding-agent performance in OpenCode than Antigravity, reinforcing that the harness is part of the model result (cheatyyyy).

DIGEST

Run /claude-api prompt-audit across existing skills to find redundant and stale rules before judging Fable 5.1 (petergyang).
Test low effort first: one report puts Fable 5.1 low at parity with Fable 5 high on CursorBench while improving cost per task (RLanceMartin).
The current migration checklist is low effort, /claude-api cost-optimize, prompt audit, mid-session effort changes, and /claude-api migrate (daniel_mac8).
For mannered output, add the literal instruction “Please remove all mannered prose” to the request or project instructions (Voxyz_ai).

DIGEST

Claude Commerce Agents is now open source, with reference implementations for retail, travel, telecom, and entertainment workflows (ClaudeDevs).
Claude Code is exploring Function Hooks for deeper extensions; this is a preview seeking feedback, not a shipped feature (ClaudeDevs).
ant apply now syncs repository declarations for managed environments, agents, skills, memory stores, and deployments to Anthropic’s API (ClaudeDevs).
/show-me is worth testing for visual PR descriptions and easier code review (mattpocockuk).
Pi v0.85.0 adds persistent per-turn Claude effort, recovery from signed-thinking mismatches, and faster transcript navigation (pidotdev).
Anthropic’s claude-code-setup plugin can inspect a project and recommend hooks, skills, MCP servers, and subagents (Hyde_ai3).
GitNexus builds a local repository knowledge graph and exposes it through MCP to Pi-adjacent tools including Codex, Cursor, and Claude Code (Hyde_ai3).