Sponsored Demo: Cut the Bill, Not the Brains. AI Coding Agents That Ship More and...
summary
Claim: routing tasks to cheaper models cuts API spend by orders of magnitude — a 12× per-token gap means 7% Anthropic usage drove over half the bill. Built an open-source Kimchi inference platform (CLI today, studio app coming) that runs Minimax/Kimian/Nematron on vLLM with prefix caching and Kubernetes-native GPU orchestration (B200 GPUs) and uses Opus only as a reviewer. Benchmarks show harness improvements lift Cloud Code from 58% to ~76.4% task success and match or beat frontier models on hard coding tasks while lowering cost.
AI-written summary · about
chapters (21)
- 0:00AI coding agents intro
- 1:02Autocomplete to autonomous agents
- 2:28Cast AI’s AI tools journey
- 3:33Kimchi inference platform
- 4:57Open-source vs frontier costs
- 6:43Harness design importance
- 7:47Stanford harness research
- 8:49Multi-agent personas workflows
- 10:13Cast AI’s GPU savings
- 10:54API pricing spiraling costs
- 11:36Agents autonomous weekends
- 12:40Kimchi studio demo
- 13:21Five key takeaways
- 15:06Q&A
- 16:08Benchmarks vs custom evals
- 16:49IDE support model selection
- 17:54Frontier models for review
- 18:58Cost quality token balance
- 19:57Harness vs token costs
- 20:58Long tasks context limits
- 22:06Summary
related talks
35mUsing coding agents with open models
59mKeynote: AI without the BS, for humans
07h 54mAIE Miami Keynote & Talks ft. OpenCode. Google Deepmind, OpenAI, and more!
34mOpen Source AI Agents on User-Owned Infra: K3s, MCP, & GPU Sharing in Practice
07h 47mAI Agents Conference 2026 - Day 5
32mRethink business value for generative AI use cases
18mSmall AI Teams with Huge Impact
08mStop Renting Your Cognitive Infrastructure