AI Digest — July 18, 2026
Quick Notes
- SQLite Query Explainer — Simon Willison had Claude Fable build an interactive tool that runs SQLite via Pyodide/WebAssembly in the browser and layers plain-language explanations onto the output of
EXPLAINandEXPLAIN QUERY PLAN, inspired by Julia Evans lamenting that she never learned to read a query plan. https://simonwillison.net/2026/Jul/18/sqlite-query-explainer/#atom-everything - Kimi K3 vs Claude Fable 5 vs GPT-5.6: The Real Cost — A head-to-head test argues Kimi K3 (Moonshot’s 2.8T-parameter open-weight model) is genuinely competitive on quality but not the budget option the charts suggest: it burned 21.5M tokens and 93 minutes on a build that Fable 5 finished in 17, so cheap per-token pricing evaporates on a cost-per-task basis. https://chaseai.io/blog/kimi-k3-vs-claude-fable-5-gpt-5-6-real-cost
Structured Summaries
New Models / Research
Moonshot AI released Kimi K3 on July 16, 2026 — a 2.8-trillion-parameter open-weight model, the largest open-weight release to date. “Open weight” here means you can inspect every tuned parameter (impossible with Fable 5 or GPT-5.6), but serving it requires millions of dollars of hardware, so most users reach it through APIs like OpenRouter. It earned headlines from two charts: Moonshot’s own benchmarks showing it trading blows with Fable 5 and GPT-5.6 across major programming tasks, and the arena.ai front-end code arena, where K3 currently ranks #1 in blind head-to-head voting.
Sources: https://chaseai.io/blog/kimi-k3-vs-claude-fable-5-gpt-5-6-real-cost
Research Insights
The core lesson from the K3 comparison is that benchmark charts measure quality and per-token sticker price, but not how many tokens a model needs to finish a job — and that gap is the whole story. On Artificial Analysis’s cost-per-task index, K3 lands within ~10% of GPT-5.6 (sometimes smaller 5.6-tier models cost half as much as K3), so the “cheap model” narrative mostly evaporates against the GPT line, though savings against Fable 5 remain real and large. K3 is also slow (≈6 min vs 5 for Fable 5 and 4.7 for GPT-5.6 on timing charts, and far worse in the author’s real front-end test). The author flags the “omniscience index” — a penalty benchmark that rewards admitting ignorance over confident wrong answers — as the metric that matters most for production agents; Fable 5 nearly doubles GPT-5.6 and more than doubles K3 there, which matters when confidently making something up is worse than stopping to ask.
Sources: https://chaseai.io/blog/kimi-k3-vs-claude-fable-5-gpt-5-6-real-cost
Tooling
Simon Willison built an interactive SQLite query explainer after Julia Evans wrote about wanting to learn to read query plans. Built by Claude Fable, the tool runs SQLite in Python via Pyodide/WebAssembly entirely in the browser and adds an explanatory layer over both EXPLAIN and EXPLAIN QUERY PLAN output. Willison notes it should be approached with caution since he can’t fully verify the explanations himself, but it “seems cromulent enough.”
Sources: https://simonwillison.net/2026/Jul/18/sqlite-query-explainer/#atom-everything
🔗 View this digest on the web: https://ainews.rusig.com/digests/2026-07-18-evening/