The AI Coding Championship Decision Kit — the Full August 2026 Scoreboard, the Decision Matrix, the Verbatim Arena Prompts, and the Exact Setup Commands for All Four Lanes
The companion to the video. Four coding agents — Claude (Fable 5), Codex (GPT-5.6 Terra), Qwen3.8-Max, and Kimi K3 — ran the same three tasks on camera: build a spec-graded app, fix a repo with three seeded bugs and 14 tests, and answer four questions about a 1,712-file monorepo with exact file-path citations. Identical prompts, fresh workspaces, sequential runs, a grader written before the first run. THE HONEST RESULT: quality tied 12/12 — every agent aced every gate — so speed and cost decided the title, and Codex (GPT-5.6 Terra) swept every round on both: 211 seconds and 32 cents sticker-equivalent for the whole gauntlet. INSIDE: (1) THE FINAL SCOREBOARD — all 12 recorded runs with quality, wall-clock, and cost per round, plus exactly how each lane's cost was counted. (2) THE DECISION MATRIX — pick Claude if you live on a Max subscription ($0 marginal, and the only agent to catch the extra comprehension nuance), Codex if you want maximum speed per dollar, Qwen if your budget is $6/mo (aced every gate, slowest overall, open weights promised but not shipped), Kimi if you need weights you can download today (on Hugging Face since Jul 27, bespoke license, not MIT). (3) THE THREE ARENA PROMPTS VERBATIM — plus the three rules that keep a rerun fair, so you can stage the same fight on your own stack. (4) SETUP COMMANDS FOR ALL FOUR LANES — Claude Code on Max, Codex CLI with gpt-5.6-terra, Qwen3.8-Max through opencode with the $6 Token Plan Lite provider block, and the Kimi K3 env swap inside Claude Code. (5) THE PRICING SHEET — every API sticker price and subscription plan verified August 4, 2026, including Terra's July 30 price cut, plus the open-weights honesty line. Independent PDF from Hyperautomation Labs — not affiliated with Anthropic, OpenAI, Alibaba, or Moonshot.
Free. No spam. Unsubscribe anytime.