started · updated
DeepSeek V4. 1 Flash nears GPT-6 Astra performance at significantly lower cost
New AI benchmarks are highlighting significant performance and cost disparities among leading large language models. In the OpenDesign Arena, which evaluates models on real-world design tasks like building web apps and dashboards, OpenAI’s GPT-6 Astra achieved the top score of 82.7 points. However, DeepSeek’s V4. 1 Flash model closely followed with a score of 81.2, while costing only $0.023 per finished design—roughly 1.4% of the $1.61 charged by GPT-6 Astra. DeepSeek achieves this efficiency by activating only 8 billion of its 552 billion parameters during prompt processing.
In a separate specialized benchmark called CAD Arena, released by Normal Research, AI models were tested on their ability to interpret technical drawings and reconstruct components within engineering environments like Siemens NX and SolidWorks. GPT-6 Astra and Anthropic’s Claude Fable 5. 1 emerged as leaders, with FrontierCAD scores of 0.671 and 0.662, respectively. The benchmark emphasizes the importance of creating models that are parametric and editable by human engineers rather than just generating static 3D shapes.
Entities
Anthropic · DeepSeek · Normal Research · OpenAI · OpenDesign Arena
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 2 SOURCES] DeepSeek V4. 1 Flash cost $0.023 per finished design, which is approximately 1.4% of GPT-6 Astra’s $1.61 cost. bitcoinethereumnews.com · decrypt.co
- [○ 1 SOURCE] In the CAD Arena benchmark, GPT-6 Astra and Claude Fable 5. 1 led with FrontierCAD scores of 0.671 and 0.662 respectively. stamparein3d.it
- [● 2 SOURCES] DeepSeek V4. 1 Flash activates only 8 billion of its 552 billion parameters to process a prompt. bitcoinethereumnews.com · decrypt.co
- [● 2 SOURCES] DeepSeek V4. 1 Flash scored 81.2 out of 100 on real-world design tasks in the OpenDesign Arena benchmark. bitcoinethereumnews.com · decrypt.co
- [● 2 SOURCES] OpenAI’s GPT-6 Astra achieved the highest score of 82.7 points in the OpenDesign Arena benchmark. bitcoinethereumnews.com · decrypt.co