Issue · 2026-09-24

Most AI news is noise.Here is the rest.

MODELS

Opus 5.5 shipped: 40% cheaper, same cost per task

Anthropic released Opus 5.5 at Fable 5.1 level for most tasks and 40% cheaper to run than Opus 5 (claudeai).
Measured: tops the Artificial Analysis Intelligence Index at 58 on $4/$20 with $0.20 cache reads, and 59.6% on Terminal-Bench 4.0 is level with GPT-6 Astra (ArtificialAnlys).
Cost per task stays flat, not 20% lower: ~119k output tokens per index task against ~73k for Opus 5 (ArtificialAnlys).
Claude Code's 5-hour limits rose 20% and the lower price stretches them ~25% further; Pro, Max and Team got a one-time reset (ClaudeDevs).
Daily-driver verdicts: writes like Opus 4.6, codes like Fable 5.1 (jarredsumner); on Victor Taelin's own prompts it beats Fable 5.1, more than Anthropic claims (VictorTaelin).
The prose is what people name first: Anthropic's first release with targeted fixes for sentence clarity and dense info dumps (JacksonKernion).
In testing next: Sonnet 5.5, 1M context, 128k output, $2/$10 — aimed at GPT-6 Sol (kimmonismus).

DIGEST

Cloud sessions left research preview — Claude Code keeps running with your laptop closed, with one-time credits of $100 Pro and $250 Max (ClaudeDevs).
Projects gained local support, so threads run on your own machine (ClaudeDevs).
Plan mode may go: a shift+tab effort-level hotkey would replace it (trq212), against the argument to let users plan how they want (mattpocockuk).
After switching: pick "opus" with the latest Claude Code CLI to get 5.5 (dexhorthy), then run /claude-api migrate for application code and /claude-api prompt-audit for skills and prompts (RLanceMartin).

MODELS

OpenAI launched both at API prices 50% lower than GPT-5.6 (OpenAIDevs); Altman's framing is half the price per token and less per task (sama).
Sol $2/$10, Luna $0.10/$0.50, cache-read discount kept. Cost per task halves to $1.06 and $0.07, on slightly more output tokens than the 5.6 models they replace (ArtificialAnlys).
Sol gains 2 points on the coding index and 6 on Terminal-Bench 4.0; Luna loses 2. Hallucination falls hard (Sol 92% to 60%), mainly because Sol declines 17% of questions, costing 5 points of accuracy (ArtificialAnlys).
Simon Willison: Luna is his favourite model for product features on cost and speed, and he wrote up all three releases with per-effort comparison grids (simonw, simonw).
Launch-day split: Sol disappointing, Luna a real surprise (Layton_Gott); Theo's summary is that Anthropic has no usable small model and OpenAI no usable large one (theo).

ORCHESTRATION

Grok Build spent weeks on the 4.7 harness: less slop, better self-validation, improved thinking efficiency and long-horizon monitoring (yunta_tsai).
Live on the API, Cursor and Grok Build (veggie_eric), with a counter-verdict that it is a disappointing release (theo).
Same task in kilocode: Grok $3.52, Opus 5.5 $7.35 (coldopn).

ORCHESTRATION

Vercel's split: an agent is brain, hands and files; separating them makes cloud runs cheap and auditable, and storage is now decoupled so a nightly memory job reads files without booting the agent (rauchg).
A copyable Cursor prompt to cut price-weighted token cost per completed task — measure the baseline first, since shrinking requests while adding turns costs more (ericzakariasson).
Anthropic's system card is the first from a lab to report multi-agent scaling to 100 parallel agents (scaling01).
Unreal Agent: open-source harness, 39% cheaper than Codex+Astra on Terminal-Bench 4.0 at equal performance (unreallabsai).
The chief-of-staff pattern: the orchestrator never touches the code, sub-agents get clean contexts, and a failed sub-agent is the retry unit instead of the whole task (arpit_bhayani).

DIGEST

/claude-api prompt-audit checks skills, AGENTS.md, CLAUDE.md and prompts, and removes anti-patterns that hobble frontier models (RLanceMartin).
One report: 140 findings, 38 of them high, across 7 repos, each tied to its patch file (gateless_zen).
It flags older-model instructions as a diff without API use (thelastminute); stale rules get ignored rather than erroring (ncsuian), so babysitter rules keep taxing Opus 5.5 silently (creedants).
For comparison: the full Opus 5.5 system prompt with tools is out at over 1.9M characters (elder_plinius), GPT-6 Sol Codex at 294k (elder_plinius).

TOOLS

Pi: pi update --models pulls the new models in (pidotdev), they are live on Radius too (pidotdev), and the new pi-voice extension adds voice (pidotdev).
Cloudflare open-sourced a skill that turns a coding agent into a security auditor; 2,428 stars in one day (RoundtableSpace).
Session hygiene: ~/.codex held 166 GB, over 100 GB of it cache, 115 GB written by two sessions this month (ziwenxu_).
Gemini 3.8 TTS is cheap and generates conversations between multiple voices (simonw).