Source details
- Original source
- The Rundown AI
- Published
- 2026-07-07
- Primary topic
- Foundation Models
Why it matters
Model launches, benchmark jumps, API upgrades, context window changes, and frontier LLM competition. This item originated as a short-form social post, so the context blocks below help expand it into tools, models, and evaluation guides.
What happened
DoorDash just published new AI research. Yes, the delivery app. Its AI research team built an internal test: 105 past code changes from its own engineers, re-run through AI code reviewers to see how many real problems each one catches. The reviewer works as a pair: one model skims every code change and flags suspicious spots, a second digs into each flag. Today, both jobs go to Anthropic models. But that test found a cheaper split: hand the skimming to Kimi K2.6, an open Chinese model, and save Claude Fable 5 for the hard part. The K2.6 / Fable split scored 65.2% at catching real problems (worse bugs count for more) vs. 53.6% for the all-Anthropic pair. It caught 8/10 critical bugs, and cost $3.81 per code change instead of $3.91. Co-founder Andy Fang says the test, called DashBench, is what lets them bring in open models. Fang: "Better quality, cheaper cost."
What to do next
Compare the hosted model pages first, then check the related tools and buyer guides before changing workflow standards.
DoorDash just published new AI research. Yes, the delivery app. Its AI research team built an internal test: 105 past code changes from its own engineers, re-run through AI code reviewers to see how many real problems each one catches. The reviewer works as a pair: one model skims every code change and flags suspicious spots, a second digs into each flag. Today, both jobs go to Anthropic models. But that test found a cheaper split: hand the skimming to Kimi K2.6, an open Chinese model, and save Claude Fable 5 for the hard part. The K2.6 / Fable split scored 65.2% at catching real problems (worse bugs count for more) vs. 53.6% for the all-Anthropic pair. It caught 8/10 critical bugs, and cost $3.81 per code change instead of $3.91. Co-founder Andy Fang says the test, called DashBench, is what lets them bring in open models. Fang: "Better quality, cheaper cost."
This AimostAll brief summarizes the linked source so readers can scan AI developments quickly and jump to the original reporting when needed.
