ํ”ผ๋“œ๋กœ ๋Œ์•„๊ฐ€๊ธฐ
๐Ÿš€ I Ran Claude Code on Every New Claude Model. Here's What Actually Ships.
Dev.toDev.to
AI/ML

๋ชจ๋ธ Routing ์ตœ์ ํ™”๋ฅผ ํ†ตํ•œ 5,000๋งŒ ๋ผ์ธ ์ฝ”๋“œ ๋งˆ์ด๊ทธ๋ ˆ์ด์…˜ 1์ผ ๋‹ฌ์„ฑ

๐Ÿš€ I Ran Claude Code on Every New Claude Model. Here's What Actually Ships.

Suraj Khaitan2026๋…„ 6์›” 20์ผ18๋ถ„advanced

Context

๋‹จ์ผ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ AI ์ฝ”๋”ฉ ๋„๊ตฌ์˜ ๋น„์šฉ ํšจ์œจ์„ฑ๊ณผ ์ฒ˜๋ฆฌ ๋Šฅ๋ ฅ ํ•œ๊ณ„ ์ง๋ฉด. ์ž‘์—… ๋ณต์žก๋„์™€ ์œ„ํ—˜๋„์— ์ƒ๊ด€์—†์ด ๋™์ผ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐœ์ƒํ•˜๋Š” ํ† ํฐ ๋‚ญ๋น„ ๋ฐ ์ถ”๋ก  ์˜ค๋ฅ˜ ๋ฐœ์ƒ.

Technical Solution

  • ๋ชจ๋ธ Capability ๊ธฐ๋ฐ˜์˜ ๊ณ„์ธต์  Ladder ๊ตฌ์กฐ(Haiku โ†’ Sonnet โ†’ Opus โ†’ Fable โ†’ Mythos) ์„ค๊ณ„
  • ์ž‘์—…์˜ Long-horizon ์„ฑ๊ฒฉ๊ณผ Risk ์ˆ˜์ค€์— ๋”ฐ๋ฅธ Dynamic Routing ๋กœ์ง ์ ์šฉ
  • Sonnet 4.6์„ Default Workhorse๋กœ ๋ฐฐ์น˜ํ•˜์—ฌ 1M Context Window์™€ ๋น„์šฉ ํšจ์œจ์„ฑ ํ™•๋ณด
  • Architecture ์„ค๊ณ„ ๋ฐ Refactoring ๋“ฑ ๊ณ ๋„์˜ Judgment๊ฐ€ ํ•„์š”ํ•œ ์ž‘์—…์— Opus 4.8 ํ• ๋‹น
  • Multi-hour Migration ๋ฐ Frontier Reasoning ์ž‘์—…์— Fable 5๋ฅผ ๋ฐฐ์น˜ํ•˜์—ฌ ์ž์œจ์„ฑ ๊ทน๋Œ€ํ™”
  • MCP(Model Context Protocol)๋ฅผ ํ†ตํ•œ ์ด์ข… ๋ชจ๋ธ ๊ฐ„ Adversarial Review ์ฒด๊ณ„ ๊ตฌ์ถ•์œผ๋กœ ์‹ ๋ขฐ์„ฑ ํ™•๋ณด

- [ ] ๊ธฐ๋ณธ ์ž‘์—…์€ Sonnet

4.6์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๊ณ  ์ž„๊ณ„์น˜ ์ดˆ๊ณผ ์‹œ ๋‹จ๊ณ„์ ์œผ๋กœ ์ƒ์œ„ ๋ชจ๋ธ๋กœ Escalation ํ•˜๋Š” ๊ตฌ์กฐ์ธ๊ฐ€? - [ ] ๊ณ ์œ„ํ—˜ Diff์— ๋Œ€ํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ์ด ์ƒํ˜ธ ๊ฒ€์ฆํ•˜๋Š” Adversarial Review ํ”„๋กœ์„ธ์Šค๊ฐ€ ํฌํ•จ๋˜์—ˆ๋Š”๊ฐ€? - [ ] ๋ฐ˜๋ณต์ ์œผ๋กœ ์„ฑ๊ณตํ•œ ๋ชจ๋ธ-์›Œํฌํ”Œ๋กœ์šฐ ์กฐํ•ฉ์„ Routine์œผ๋กœ ์ž๋™ํ™”ํ•˜์—ฌ ์Šค์ผ€์ค„๋งํ•˜๊ณ  ์žˆ๋Š”๊ฐ€? - [ ] ์ตœ์ƒ์œ„ ๋ชจ๋ธ์˜ ๊ฐ€์šฉ์„ฑ ๋ณ€๋™์— ๋Œ€๋น„ํ•œ Fallback ๊ฒฝ๋กœ๊ฐ€ ์„ค๊ณ„๋˜์–ด ์žˆ๋Š”๊ฐ€?

์›๋ฌธ ์ฝ๊ธฐ