DeepSeek V4.1-Flash-1017: vendor benchmarks and ~$0.20 per task
The author walks through DeepSeek's model card: 763B parameters, 1M context, MIT weights, Terminal-Bench 4.0 up from 31.2 to 42.3, DeepSWE at about $0.20 per task versus $0.50 before, and Search/Fetch APIs at $0.75 per 1,000 calls.
Can a "minor" model update really matter? DeepSeek's V4.1-Flash-1017, released October 17, makes a decent case. DeepSeek describes it as a minor update to V4.1-Flash, and the weights are on Hugging Face under the MIT license (763B parameters, 1M-token context). On DeepSeek's model card, every listed benchmark improves over the previous Flash, with Terminal-Bench 4.0 going from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0. The part I find most useful is cost. In DeepSeek's DeepSWE chart, as reported by The Decoder, it averages just under $0.20 per task, compared with around $0.50 for the predecessor, helped by shorter outputs. DeepSeek also opened Search and Fetch APIs at $0.75 per 1,000 calls. These are vendor numbers, so test them on your own work. If you're choosing a model for an agent, take ten real tasks from your backlog, run them on your current model and on this one, and log success, time, and cost for each. Keep that table and rerun it every time a new model ships. 🧠 #MachineLearning Link in Comments