The Eyes Kit — Give the Claude Code You Already Use Eyes with GLM-5.3-Flash on the $18 Plan: the Exact 3-Line Config + the [1m] Context Fix, My Screenshot-Debugging Prompts Verbatim, the Rerunnable App Battery, and the Pricing Sheet with the Sep 9 Cliff
The companion to the video. On August 26, 2026 Z.ai unmasked the mystery model 'Ox Alpha' as GLM-5.3-Flash — the first natively multimodal GLM: 320B parameters (18B active), a 1M-token context window, MIT weights on Hugging Face, and roughly 9×18× cheaper than regular GLM-5.3. I verified the part that matters the week it dropped: it works inside Claude Code on the $18/month GLM Coding Plan INCLUDING vision — you paste a screenshot instead of writing a bug report. PAGE 2 — THE CONFIG: the exact three-line block (ANTHROPIC_BASE_URL / ANTHROPIC_AUTH_TOKEN / ANTHROPIC_MODEL='glm-5.3-flash[1m]'), why the [1m] suffix matters (Claude Code doesn't recognize the model yet and shrinks the window without it), how to hand it a screenshot, and the weekend half-rate billing trick. PAGE 3 — THE PROMPTS: my screenshot-debugging prompts word for word — fix-from-screenshot-only, debug-from-user-report, and build-from-mockup. PAGE 4 — THE BATTERY: rerun my four day-one tests on your own machine — the 2-planted-bug vision fix (2/2 in 80s), the failing-suite bugfix (5/5 in 20s), the one-shot expense-splitter app (gorgeous, and my click-through test found 3 real bugs), and the closer: it fixed its own app from a screenshot in 3m27s and flagged the other two bugs unprompted. Stock Claude Code control runs included. PAGE 5 — THE MONEY PAGE: the full pricing sheet (promo $0.075/$0.25 per million ends September 9, 24:00 UTC+8; list $0.15/$0.50; regular GLM-5.3 $1.40/$4.40) and the three honest catches, including the one that matters for client work: your code travels to z.ai servers in China. Every claim verified live on my machine August 29, 2026. Independent kit from Hyperautomation Labs — not affiliated with Z.ai or Anthropic.
Free. No spam. Unsubscribe anytime.