Cursor Adds DeepSeek V4.1 Flash to Its IDE After Model Sets Sub Dollar Cost Frontier on CursorBench 4.0
DeepSeek V4.1 Flash is now available in Cursor! Every effort level sits on the CursorBench cost frontier: from $0.60 to $0.98 per task, no model scores higher. Turn it on in Cursor Settings > Models.
Context
Delivering a 47.7 percent CursorBench 4.0 score at 0.60 to 0.98 dollars per task gives software teams a frontier coding option that costs a fraction of flagship proprietary models while outperforming GPT 5.6 Sol and Grok 4.7 at comparable price points.
Cursor regularly benchmarks frontier and open weight coding models on CursorBench 4.0 to map the tradeoff between task completion accuracy and average dollar cost per software engineering task. Earlier on October 26, Cursor also introduced its Agent SDK for defining and deploying autonomous coding agents as versioned Markdown and TypeScript code.
Deep dive
On October 26, Cursor announced that DeepSeek V4.1 Flash is now available to developers inside the Cursor editor under model settings. According to the accompanying CursorBench 4.0 evaluation chart comparing model score against average cost per task, every reasoning effort level of the DeepSeek V4.1 Flash 1017 checkpoint sits directly on the sub dollar cost efficiency frontier.
On the CursorBench 4.0 chart, DeepSeek V4.1 Flash 1017 clusters between 0.60 dollars and 0.98 dollars per task and reaches a peak accuracy score of 47.7 percent at 0.98 dollars per task. At that price tier, no other evaluated model scores higher: GPT 5.6 Sol drops to roughly 24 percent accuracy near the 0.98 dollar mark, Grok 4.7 falls from 46 percent at 6 dollars down to roughly 33 percent at 1.50 dollars, and Claude Haiku 5.5 drops steeply below 40 percent near 0.50 dollars.
Higher scoring proprietary models on the chart require substantially greater spend per task, with Claude Sonnet 5.5 peaking near 55 percent between 3 dollars and 9 dollars per task, Gemini 4 Argon clustering around 50 to 53 percent at 4 to 5 dollars, Fable 5.1 holding near 52 percent between 12 and 18 dollars, and Claude Opus 5.5 reaching roughly 57 percent around 13 dollars per task.
What to expect next
By establishing an uncontested Pareto frontier below one dollar per task on realistic repository benchmarks, DeepSeek V4.1 Flash makes continuous background agent loops economically viable for routine refactoring, test generation, and pull request triage. Expect AI coding platforms to increasingly route high volume background subtasks to sub dollar open weight Mixture of Experts checkpoints while reserving ten dollar frontier reasoning runs for complex architectural changes.