The open-source 'Gemma 4 31B' can achieve the same quality as the 'Claude Sonnet 5' for some tasks at 1/40th of the cost.



New AI models keep emerging and breaking benchmark score records, but for many tasks, a less powerful AI is often sufficient.

AlphaSense , a company that develops AI-powered document management services, has published test results showing that its open model, Gemma 4 31B, performs comparably to Claude Sonnet 5 in financial tasks.

How AlphaSense Pairs Frontier AI Models with Frontier Context
https://www.alpha-sense.com/resources/product-articles/frontier-ai-models-context/

When using AI to analyze financial information, the method of feeding the information significantly impacts task efficiency. A method that involves feeding all relevant information into the AI results in a large amount of data processing, increasing the cost per task. On the other hand, a method that extracts specific information from related data using a separate search system and then feeds it into the AI can reduce costs. AlphaSense has developed its own information retrieval platform, 'AlphaSense Search,' and has measured and published the financial information analysis performance when combining AlphaSense Search with various AI models.

The following are the results of measuring the accuracy of 245 financial information analysis tasks using 'GPT-5.6 Sol,' 'Claude Haike 4.5,' 'Claude Sonnet 5,' 'Claude Opus 4.8,' 'Claude Opus 5,' 'Kimi K3,' 'GLM-5.2,' 'Inkling,' and 'Gemma 4 31B.' The vertical axis shows the relative score of the accuracy (higher is better), and the horizontal axis shows the median cost for each task (lower is better). Gemma 4 31B is an open model that is freely available, but it succeeded in processing with the same accuracy as the closed model Claude Sonnet 5, at 1/40th the cost. The most accurate was GPT-5.6 Sol, which was evaluated as a 'model with an excellent balance of quality and price.'



Google's official X account also responded to the AlphaSense test results, stating, 'The most advanced model isn't always necessary. In recent benchmarks, the Gemma 4 31B achieved the same response quality as the Claude Sonnet 5 at about 1/40th the cost. Its cost-effectiveness and low latency allow Gemma to handle high-volume use cases where larger models are not economically viable,' highlighting the advantages of the Gemma 4 series.




Related Posts:

in AI, Posted by log1o_hf