Co-Design Bench is out. Built by researchers at Alibaba and several Asian universities, it gives coding agents a harder, more realistic test: editing existing web apps, not spinning up polished demos from scratch.
In the paper, the team says it started with 37,463 problematic real user requests and narrowed them to 273 tasks. Those tasks include multimodal inputs such as images, receipts, audio, video, tables, and documents, and they span everything from small layout or data edits to full project overhauls.
In Co-Design Bench’s 10-model test, GPT-6 Astra came in first, but even then it fully completed 26.4% of tasks and met 69.9% of the criteria on average, according to the researchers. Claude Opus 5 and Qwen3.8-Max followed behind it. Grok 4.6 landed at the bottom. One wrinkle: Qwen3.8-27B placed second on the larger overhaul tasks, though the authors stop short of saying smaller models are better in general.
If you use coding agents, this is one to watch. The paper says every model struggled when a request mixed layout, design, interaction, and content changes in the same job. Older coding benchmarks are already saturated above 99%. Co-Design Bench still leaves a gap of more than 35 points, using strict browser-based grading and a second judge that checks whether an agent leaned on mock-ups or hardcoded outputs.
The benchmark also tracks with SWE-bench Verified results, the researchers say: GPT-4o at 30.7%, GPT-4.1 at 54.6%, and Qwen3-Coder at 67.4%. That lines up with the broader point here. These tools still need debugging and review from you.
The paper and the benchmark are available online from the researchers.
Author: Jesús Pina
{
"social": {
"email": "",
"facebook": "",
"twitter": "",
"linkedin": ""
},
"ja-JP": "",
"de-DE": "Jesús Pina verfolgt, was künstliche Intelligenz an der Software wirklich verändert, die wir täglich benutzen — und trennt dabei, was schon funktioniert, von dem, was bisher nur eine Ankündigung ist. Ihn interessiert die praktische Wirkung: welches Werkzeug echte Zeit spart, welches auf halbem Weg stehen bleibt und welches man sich vor der Installation zweimal ansehen sollte. Er arbeitet täglich mit KI-Modellen mitten im redaktionellen Prozess von Softonic und kennt deren Grenzen deshalb recht genau.",
"en-US": "Jesús Pina follows what artificial intelligence is actually changing in the software people use every day, telling apart what already works from what is still only an announcement. What interests him is the practical effect: which tool saves real time, which one stops halfway and which one deserves a second look before you install it. He works daily with AI models inside Softonic's own editorial process, so he knows their limits fairly well.",
"es-ES": "Jesús Pina sigue lo que la inteligencia artificial está cambiando de verdad en el software que usamos cada día, separando lo que ya funciona de lo que por ahora solo es un anuncio. Le interesa el efecto práctico: qué herramienta ahorra tiempo real, cuál se queda a medias y cuál conviene mirar dos veces antes de instalarla. Trabaja a diario con modelos de IA dentro del propio proceso editorial de Softonic, así que conoce bastante bien sus límites.",
"fr-FR": "Jesús Pina suit ce que l'intelligence artificielle change réellement dans les logiciels du quotidien, en distinguant ce qui fonctionne déjà de ce qui n'est encore qu'une annonce. Ce qui l'intéresse, c'est l'effet pratique : quel outil fait gagner du temps, lequel s'arrête à mi-chemin et lequel mérite un second examen avant installation. Il travaille au quotidien avec des modèles d'IA au cœur du processus éditorial de Softonic, et connaît donc assez bien leurs limites.",
"it-IT": "Jesús Pina segue ciò che l'intelligenza artificiale sta davvero cambiando nel software di ogni giorno, distinguendo quello che già funziona da quello che per ora è solo un annuncio. Gli interessa l'effetto pratico: quale strumento fa risparmiare tempo, quale si ferma a metà e quale conviene guardare due volte prima di installarlo. Lavora ogni giorno con modelli di IA dentro il processo editoriale di Softonic, e per questo ne conosce bene i limiti.",
"nl-NL": "Jesús Pina volgt wat kunstmatige intelligentie echt verandert aan de software die we dagelijks gebruiken, en scheidt daarbij wat al werkt van wat nog slechts een aankondiging is. Wat hem interesseert is het praktische effect: welke tool echt tijd bespaart, welke halverwege blijft steken en welke je beter twee keer bekijkt voordat je hem installeert. Hij werkt dagelijks met AI-modellen binnen het redactieproces van Softonic en kent daardoor hun grenzen vrij goed.",
"pl-PL": "Jesús Pina śledzi to, co sztuczna inteligencja rzeczywiście zmienia w programach używanych codziennie, oddzielając rozwiązania, które już działają, od tych, które są na razie tylko zapowiedzią. Interesuje go efekt praktyczny: które narzędzie naprawdę oszczędza czas, które zatrzymuje się w połowie drogi, a które warto sprawdzić dwa razy przed instalacją. Codziennie pracuje z modelami AI wewnątrz procesu redakcyjnego Softonica, więc dobrze zna ich ograniczenia.",
"pt-BR": "Jesús Pina acompanha o que a inteligência artificial está realmente mudando nos programas que usamos todos os dias, separando o que já funciona do que ainda é apenas anúncio. O que lhe interessa é o efeito prático: qual ferramenta economiza tempo de verdade, qual para no meio do caminho e qual merece uma segunda olhada antes de instalar. Trabalha diariamente com modelos de IA dentro do próprio processo editorial da Softonic, e por isso conhece bem seus limites."
}
View all posts by Jesús Pina
