Use DeepSeek V4 Pro in Cursor: Complete Setup

Cursor runs on OpenAI-compatible providers, and DeepSeek exposes exactly that contract at https://api.deepseek.com^[1]. With the 0813 GA, V4 Pro became a credible in-IDE coding model: 96.40% on SWE-bench Verified (third-party neutral harness) at $0.022 per test^[5], with MIT-licensed weights behind it.

The August 13 GA also shipped native Responses API support and a Codex integration path^[2][4], which matters for Cursor because Cursor's agent uses OpenAI-style tool calling under the hood — the same contract DeepSeek implements.

• One base_url, OpenAI SDK, tool calls, JSON output — everything Cursor needs.
• Reasoning effort low/high/max for fine-grained cost control^[1].
• 1M-token context window fits large-repo agent loops^[6].

In Cursor, open Settings → Models → Add a new OpenAI-compatible provider and point it at DeepSeek. The official quick-start documents the same parameters you will reuse everywhere: base_url https://api.deepseek.com and your DeepSeek API key^[1].

• Base URL is https://api.deepseek.com — do not append /v1 (DeepSeek accepts it, but the documented root is cleaner).
• Use the same key as the official API; there is no separate Cursor key.
• If your Cursor build needs /v1, it works too — DeepSeek is prefix-agnostic^[1].

# Cursor → Settings → Models → "Add Provider"
# Provider type: OpenAI-compatible
Name:        DeepSeek
Base URL:    https://api.deepseek.com
API key:     <DeepSeek API Key from platform.deepseek.com>

# Then enable these model IDs in the Models list:
#   deepseek-v4-pro
#   deepseek-v4-flash

Model names are stable: deepseek-v4-pro resolves to the 0813 GA checkpoint, deepseek-v4-flash to the 0731 build^[1]. Cursor lets you set per-model parameters; pass reasoning_effort through extra-body configuration where your Cursor version supports it^[1][3].

DeepSeek's official announcement frames effort levels as: low for simple tasks, high for daily agent workflows, max for complex tasks^[2]. In Cursor, match effort to what the tab is doing — low during autocomplete-style edits, high or max for refactors and architecture changes.

DeepSeek returns chain-of-thought in the reasoning_content field of the OpenAI-format response^[3]. Some Cursor builds forward only the content field and drop reasoning_content, which can truncate long thinking traces or break streaming mid-thought on older versions.

If you hit odd truncation, the community workaround is a small local proxy that rewrites reasoning_content into the content stream before it reaches Cursor (the deepseek-cursor-proxy pattern). This is a model-access fix, not a harness integration, and is optional — most users run fine without it after the 0813 update.

The proxy pattern is simple: a Python/Node server listens on localhost, forwards requests to DeepSeek with the reasoning_content field, and merges it into the visible message so Cursor's UI doesn't drop the thinking trace. You keep your API key local, and the only downside is an extra hop that adds a few milliseconds. If you see 'response ended mid-thought' or empty tool-call reasoning, this is the first thing to try.

A cleaner alternative: use the official Codex path — DeepSeek's one-click Codex setup configures models.json with correct reasoning metadata for agent clients^[4]. Cursor's agent mode benefits from the same metadata being present.

For Cursor's agent/composer mode, treat V4 Pro like a frontier coding model with a budget: keep the shared system prompt and repo context stable so prefix caching kicks in (cache-hit input drops to $0.022 off-peak after the 8/16 pricing change)^[1].

The combination of MIT weights, native tool calling, and sub-$0.02-per-test coding accuracy makes V4 Pro a strong default inside Cursor for open-source teams — with the closed flagships still ahead only on long-horizon autonomous tasks^[5][7].

Budget tip after the 8/16 pricing change: a full day of Cursor agent sessions on V4 Pro at off-peak rates costs about a third of what the same sessions cost at peak. If your team is in a timezone where your working hours overlap the UTC peak windows (01:00-04:00 and 06:00-10:00 UTC), consider shifting heavy agent runs or accepting the peak price for interactive work — the difference is roughly 2x on every token^[1].

• Keep a consistent @-mentioned file set per task to maximize cache hits.
• Schedule long agent runs into off-peak hours (all hours except 01:00-04:00 and 06:00-10:00 UTC) to halve the new rates^[1].
• Escalate from V4 Flash to V4 Pro only when a task needs deep reasoning — the hybrid pattern cuts cost ~3x^[6].
• Test the tool-call loop on a small repo first; tool calling is fully supported but tuning effort/thinking per step helps.