
ALT Card 1 of 2. Poetry, planets and the periodic table: AI model families, November 2026, API list price per million tokens in and out. Claude, poems then stories: Haiku $0.10/$0.50, Sonnet $2/$10, Opus $4/$20, Fable $8/$40, Mythos $8/$40 (vetted). OpenAI, moon, Earth, Sun, stars: Luna $0.10/$0.50, Terra $1/$5, Sol $2/$10, Astra $8/$40, GPT-6 cyber (vetted, trusted access). Gemini, elements out of order: Hydrogen $1.50/$7.50, Argon and Carbon $2/$10 rising to $4/$20, Neon $3/$15 (new), Carbon for cyber (vetted, Fairwind). Haiku and Luna cost the same, so do Sonnet and Sol, and Fable and Astra.

ALT Card 2 of 2. Cost vs capability on the benchmarks: AI model families, November 2026, as a scatter chart. Across: price per million output tokens, log scale. Up: Artificial Analysis Intelligence Index. Claude (orange): Haiku $0.50, 36; Sonnet $10, 55; Opus $20, 59; Fable $40, 65. OpenAI (blue): Luna $0.50, 28; Terra $5, 51; Sol $10, 57; Astra $40, 62. Gemini (green): Hydrogen $7.50, 54; Argon $10, 50; Carbon $10, 60; Neon $15, 61 (Google's launch chart said 67). Callouts: Haiku and Luna are the same price, 8 points apart. Opus: mid-table on tests, builds our code every day. Argon and Carbon go to $20 on 1 December.

ALT The bottleneck has moved. Full suite on a real database: 64 to 81 min: A unit-test job runs about 25 min against a 30 min cap. It got cancelled twice in one day on a slow runner. Merge only on the tip of main: Every merge makes every other open PR stale. That's roughly one change an hour, however fast the agents write. So we batch changes into tested trains: One full run proves several changes at once. How we pick what's safe to batch stays in-house. A script merges, never the author: Then a live health check, then someone sees it working in production.

ALT AI is only as good as its access. Licence: A Copilot or Claude seat and a chat window. A great start. Access: Email, code, the project tracker. Where my biggest productivity jump came from. Trust: Deciding what it's allowed to read and do. Governance: Permissions you can explain to someone else.

ALT Verify the behaviour, not the proxy. Green CI is a proxy: Every test was green. The real export produced 0 of 5 review PDFs. Walk it in a real browser: The way a user would, in the real environment. A quiet log proves nothing yet: It counts once you've seen it record a known good case. End every task with this footer: Read the Unverified line first. Verified: what behaviour, where, what evidence Unverified: what I didn't check Residual risk: HIGH / MED / LOW If the gap hides a bug, the symptom would be: …























