GPT-6 Astra and Claude Fable 5.2 hit a 5,384-task bloodbath and every AI founder is…
This is a dev post classified by Jev as AI dev tools (a tutorial), kept by the Dev Radar because it carries real work, not commentary.
GPT-6 Astra and Claude Fable 5.2 hit a 5,384-task bloodbath and every AI founder is sweating Microsoft Research released LoopsBench across 112 long-horizon software projects in 8 programming languages, evaluating persistent agents across 5,384 separately testable units: Claude Opus paired with an outer continuation loop topped the leaderboard, resolving 25.00% of tasks while burning thousands of tokens per run Most enterprise teams evaluate models on localized single-shot benchmarks like SWE-bench, expecting prompt tweaks to fix production code. LoopsBench shows that single-turn harnesses fa
Posted by marfin (3.5k followers) 1 h ago · 6 likes · 342 views · view the original post on X. Kept by the Dev Radar as AI dev tools.
More dev work like this
- we did this not because we thought it would be easy. we did this because we wanted to… — @michellechen
- 让 Agent 操作浏览器实现自动化，经常会遇到各种问题，每个站都得重复登录，若碰到人机验证基本卡死。 — @GitHub_Daily
- …is this what it takes to solve "continual learning"? — @teortaxesTex
- Grok Build v1.0.40 is out with a few practical updates. The main item is miscellaneous… — @blankspeaker
- We're hosting a hot topic livestream on Jev with — @LangChain
- Browser automation gets harder when your agent starts from a blank, logged-out profile. — @DanKornas
- One day after Laya and 4 after Jev, @cognition releases Kev another open model that… — @plotarmordev
- Indexes your codebase, dependency graph, and git history to give AI agents cited answers… — @tom_doerr
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 17k posts from 4.9k X accounts over the last 21 days, 1.9k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-21 00:32 UTC. Full method.