GPT‑5.6 is helping Vercel’s internal data science agent tackle harder work Andrew Qu, Vercel’s Chief of Software, used GPT-5.6 and Codex to unblock an internal data science agent on work where earlier models got stuck.

Build with GPT-5.6:

@vercel @andrewqu Return these excellent models. #Keep4o #Keep51 #Keep45 #Keep41 #keepo3

@vercel @andrewqu what a cutie

@vercel @andrewqu 🐐

@vercel @andrewqu 他不是甄子丹吗？

@vercel @andrewqu unblock is the key word here. getting past that local minimum is usually the hardest part of agent workflows.

@vercel @andrewqu Cringe

@vercel @andrewqu Lfg!

@vercel @andrewqu 5.6 doesn't have any improvements in coding and cyber security. Pretty disappointing modal so far.

@vercel @andrewqu It is helping not only Vercel! It is helping all of us 🔥

@vercel @andrewqu I met and talked to @andrewqu in person once. He remembered my profile picture. Absolute goated 🐐

@vercel @andrewqu @LilysAI_要約して

@vercel @andrewqu lfg @andrewqu @danizeres !!! killin it!

@vercel @andrewqu Codex is the real hero here.

@vercel is the agent handling 600+ concurrent tasks or still batch-queued

@vercel so which model was the bottleneck - 4o or o3

@vercel Well said!

@vercel curious what the data science agent was trying to do

@vercel what’s the hardest thing it actually unblocked - a query they couldn’t optimize or a data pipeline that kept timing out

@vercel @andrewqu codex in a data agent eval gates unblocked the flow

@vercel wait what was it getting stuck on before

@vercel was it a reasoning bottleneck or a context length issue?

@vercel so codex is the real mvp here or is gpt-5.6 actually better at chaining logic

@vercel @andrewqu guy in vid is cool

@vercel agent loop or model - which was the real bottleneck

@vercel when’s the public release for that agent capability?

@vercel fr what tasks were they stuck on

@vercel was it the reasoning loop or the tool use that finally clicked for the agent 💀

@vercel gayyyyyyyyyy

@vercel tried it on a messy airflow pipeline yesterday it actually figured out the dag dependencies where 4o just gave up

@vercel wait what kind of data science agent? codex or a custom pipeline?

@vercel what’s the real success rate bump vs 4o on that agent

@vercel Thanks for sharing this

@vercel Very meaningful.