Seven chapters on where AI actually stands in mid-2026 — written to be updated, not archived. The model comparisons go stale fast. This page is the living version; the PDFs are rendered from it.
OpenAI's GPT-5.6 family, tier by tier. Why Terra — not the flagship — should be your default, the benchmark OpenAI attacked the day before launch, and the 1M-token context window that Luna cannot actually use.
↓ Download PDF · 9 pagesEvery model that matters: Claude, GPT, Gemini, DeepSeek, Kimi, Qwen, Grok, and the open-weight insurgency now processing the majority of tokens on OpenRouter. Plus the arithmetic on why self-hosting is almost always a mistake.
↓ Download PDF · 9 pagesAnthropic and OpenAI head to head, with the criticism of both left in. 65% of developers prefer Codex day-to-day; 67% rate Claude's code better in blind review. Both are true.
↓ Download PDF · 8 pagesAntigravity isn't a model — it's Claude Code's most credible challenger. Plus Codex, Cursor, Copilot, and the Claude Code features that shipped while you weren't looking.
↓ Download PDF · 7 pagesHow the people who build Claude actually use it. The techniques that are now no-ops, the ones that are actively harmful, and the design bias in Opus 4.8 that has been quietly fighting you.
↓ Download PDF · 11 pagesThe real differences between the five surfaces Claude Code runs on — and what Cowork is, now that it's a third tab you have never clicked.
↓ Download PDF · 5 pagesHow to update this. Edit the chapter files in ~/Desktop/AI-Field-Guide/chapters/, then run ./build.sh to re-render every PDF into ~/Desktop/Education, and ./deploy.sh to push the site. Update once — the page and the PDFs stay in sync.
Every figure is sourced. Anything that could not be verified is marked as such rather than guessed, and where the two labs' own benchmark claims conflict, both are shown.