Los Angeles – September 15, 2026 -- Alibaba.com's Accio AI agent platform completed a 107-task e-commerce benchmark for an estimated $3.69, more than 50% cheaper than OpenAI's Codex at $9.27 and Anthropic's Claude Code at $9.51, according to results the company disclosed at its CoCreate event for entrepreneurs and small businesses.
Benchmark shows Accio undercuts general-purpose AI agents by over 50% on cost
The evaluation, run across 107 real-world commerce tasks, found Accio matched the completion quality of the two rival systems while sharply reducing computing spend. Alibaba.com President Kuo Zhang said unaffordable AI is useless for small businesses, framing the company's objective as making commerce AI practical for a one-person operation rather than simply more powerful.
Task-specific model routing drives the cost reduction
Accio's savings stem from post-training lightweight models on commerce-specific data for routine tasks, reserving more capable models for complex reasoning. The system breaks requests into steps and allocates resources per step to balance quality, speed, cost and data needs, rather than defaulting to either the cheapest or most powerful model available. Cache reuse, context compression and coordinated agent execution further cut repeated processing and redundant token consumption.
Accio expands into a unified workspace linking five major storefront platforms
Alibaba.com said Accio now functions as a single workspace covering market research, product development, supplier evaluation and daily operations for small sellers. Integrations with Amazon, Shopify, eBay, TikTok Shop and Walmart give sellers access to supported storefront workflows from one interface.
Alibaba.com open-sources Commerce Agent Bench built from 10 million SMB users
The company released Commerce Agent Bench on GitHub, built from real merchant activity rather than synthetic exercises. The dataset draws on 10 million active SMB users, 1.6 million conversations and 200,000 execution traces, distilled into 107 end-to-end tasks spanning seven categories and four levels of autonomy, including reviewing unstructured emails, detecting payment fraud, calculating landed costs and booking multi-carrier shipping routes. No single AI model led across all task categories in the benchmark, a result Alibaba.com said reinforces the case for routing different tasks to different specialized models instead of relying on one general-purpose system.