Issue #4 ·

GPT-5.6 Lands

Wednesday was chaos. OpenAI dropped the GPT-5.6 family — Luna, Terra, Sol — Meta shipped Muse Spark 1.1 with a public API, and Anthropic started charging per token for Fable 5 on Pro plans. Three companies, three moves, one signal: the model churn isn't stopping and neither is the pricing reshuffle.

The tools

Star of the week

no-mistakes coding open source

An AI review gate that runs before code reaches the shared repo.

A last line of defence for agent-written code, at the exact moment it matters.

olmocr by Allen AI research open source

Open-source document OCR and PDF understanding from Allen Institute for AI.

Clean PDF ingestion is still the Achilles heel of most RAG pipelines. This is the strongest open-source option for it.

TencentDB Agent Memory infra open source

Team-level memory hub for AI agents — turns conversations, docs, and code into shared, reusable memory assets across agents and frameworks.

Agent memory is where everyone rolls their own — a governed, shareable layer beats yet another private vector store.

claude-video image-video open source

Gives Claude the ability to watch any video — /watch downloads it, extracts frames, transcribes, and hands the lot to the model.

Video is the last big blind spot for text-first agents; this closes it with plain plumbing, no new model required.

Worth reading

Simon Willison's notes on GPT-5.6 — programmatic tool calling, multi-agent support, and three model tiers. The best 5-minute read on why this release matters for builders.

From the big labs

Meta's first paid API — $1.25/$4.25 per million tokens, $20 free credits, and the quiet end of the open-only era. A frontier-tier agentic model you can now simply buy.

xAI's first model aimed squarely at coding and agents — $2/$6 per million, trained on real Cursor sessions, pitched by Musk as comparable to Opus 4.7 but much faster.

Get the next one in your inbox

One email, every Thursday. The goodies that matter — nothing else.