How DeepSeek V4.1 Flash compared to Claude and GPT as an agent in Salesforce

I've been working on a project testing AI agents in a Salesforce dev org and wanted to share how DeepSeek V4.1 Flash did, since it was the best value of the five models I tested. Five models (DeepSeek V4.1 Flash, Claude Opus, Sonnet, Haiku and GPT-6.1 Sol) worked through 50 tasks, and each attempt was graded on what actually changed in the database, not on the model's reply. Some of the findings for DeepSeek: It scored 82%, close to Opus (87%), and the gap between them wasn't statistically significant It cost 0.11 cents per task, about 75 times less than Opus (8.4 cents) When Salesforce temporarily locked a record, it kept retrying and finished every time, while Opus gave up on most Its main weakness was rarely re-checking records after changing them, so it missed some changes Salesforce made on its own If anyone's interested: Write-up: outsideengineer.substack.com/p/evaluating-ai-agents-on-salesforce Code and transcripts: github.com/brianna-rose05/crm-agent-bench