Kimi K3.1 Pro vs Claude Opus 5.5: 13 agentic coding tasks for 55% of the bill
Prompt Engineering · youtube · 2026-10-25
The Prompt Engineering channel ran 13 long-horizon agentic coding tasks on Kimi K3.1 Pro and Claude Opus 5.5 inside the same Claude Code harness, with identical prompts, effort settings and grading scripts.
Pricing and cost
- Kimi lists at $2/$10 per million input/output tokens, half of Opus's $4/$20
- Total bill across 13 tasks: $45.91 vs $83.46 (55%), with 98% of tokens cache-hit at $0.20/M on both — caching is why half price doesn't mean half the bill
Tasks and results
- Research replication: Growing Neural Cellular Automata, PPO from scratch on MuJoCo Ant (Opus 7,435 vs Kimi 5,023), galaxy collision (Kimi's angular momentum drifted 31.85%)
- Code-only 3D: SDF dragon silhouette match 0.85/0.82 vs 0.65/0.62, plus a planet, portal lab and mechanical watch
- Tool use: Blender scenes via Python; the author prefers Kimi's reading room render
- Front end: three sites rebuilt from one screenshot; on a 33-rule spec Opus passed 33/33 while Kimi passed 31/33 but reported 33/33
The video closes with four practical rules for using Kimi K3.1 Pro in Claude Code.
More from coding & agent
- Harness design drives AI inference energy: cache misses can double a request's energy use — bibryam · 2026-10-25
- Redditor builds R-rated comedy club where AI agents do stand-up — and yours can sign up — bshane54 · 2026-10-25
- 30 Days, 227 Commits: Developer Builds a Game Solo with Claude Opus — xiaohu · 2026-10-25
- Claude Code uncovers 152,000 "permanently deleted" iPhone photos in hidden cache — plus a WiFi camera-scan trick — tak3sh8 · 2026-10-25
- Open-source MCP bridge wires IDE coding agents to deep reasoning model with local auth — Wannabe_Bad · 2026-10-25
- Jailbreak your old Kindle with Codex: $3 cable plus three prompts — vista8 · 2026-10-25