GPT 5.2 Codex's agentic benchmark performance

Hey there!

Welcome back to The Pulse, where we dive into interesting AI stories and trends backed by data, all presented through simple visuals.

> OpenAI’s most advanced agentic coding model released; available today on ChatGPT (API in coming weeks)

> major improvements on context compaction, long-horizon work & cybersecurity + SOTA on agentic coding benchmarks

> strong focus on cybersecurity: not yet “high” level in capability, but significant improvement over previous models

> OpenAI also launched own App Store - devs can submit apps to be published on ChatGPT for 900M users

> based on Cloudflare's public DNS query data

> real 2025 user traffic shows new GenAI entrants: Google Gemini, Windsurf AI, Grok & DeepSeek

> AI bots = 4.2% of global HTML requests (direct AI consumption of the web) + Googlebot additional 4.5% (dual-purpose: search indexing + AI training) = ~9% overall

> AI “user-action” crawling grew >15× in 2025: driven by live user queries inside AI chatbots (browsing, answer generation, etc.), although training activity tops

> AI crawlers most commonly blocked bots in robots.txt as website owners try to stop AI scraping

> past 2 years: increase in both wages & jobs in occupations most exposed to AI

> occupations most exposed to AI: ~140 where today’s AI can automate work hours with moderate human oversight at satisfactory quality

> study claims current AI raises productivity & moves workers toward higher-value tasks instead of replacing

> also claims AI adoption rates range from 3.6% to 25.4% today with 7.5%–15% of work hours automated by 2028 depending on industry