DeepSeek V4 Flash in Cursor, Claude Code & Codex: Setup Guide

DeepSeek V4 speaks two native agent dialects. The OpenAI-compatible endpoint at https://api.deepseek.com and the Anthropic-compatible endpoint at https://api.deepseek.com/anthropic^[4] both accept the same sk- API key. Create one key at platform.deepseek.com/api_keys and it works in every tool below.

• Create your key at platform.deepseek.com/api_keys. It starts with sk- and runs pay-as-you-go, so add balance before heavy use.
• Model names are case-sensitive. Use deepseek-v4-flash for fast, cheap, high-frequency work and deepseek-v4-pro for complex reasoning^[5].
• The old aliases deepseek-chat and deepseek-reasoner retired on July 24, 2026^[1]. Requests to them now return a hard error.
• Append [1m] to select the 1M-context variant of a model, for example deepseek-v4-pro[1m].

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -d '{
        "model": "deepseek-v4-flash",
        "messages": [
          {"role": "user", "content": "Hello!"}
        ],
        "stream": false
      }'

Adding DeepSeek V4 Flash in Cursor takes about two minutes. DeepSeek ships no dedicated Cursor integration page^[6], so this is the community-verified path: Settings to Models to Add Model.

Once Verify passes, select the new model in the Chat dropdown before opening a thread. Pro and Flash can coexist as two entries sharing the same key and base URL.

Cursor enforces its own per-plan context ceiling, so in practice you hit Cursor's limit before DeepSeek's 1M-token window.

• Open Settings (Cmd+, on macOS or Ctrl+, on Windows) and go to Models.
• Toggle Override OpenAI Base URL and enter https://api.deepseek.com. Do not append /v1.
• Paste your DeepSeek API key into the OpenAI API Key field. Cursor stores it locally and forwards it as a Bearer token.
• In Model Names, click + Add model and type the exact name: deepseek-v4-flash or deepseek-v4-pro. There is no deepseek-v4 alias.
• Click Verify to test the connection.

DeepSeek's official Claude Code integration is the most authoritative of the three^[3]. Point Claude Code at the Anthropic endpoint, set the eight environment variables, and Claude Code routes to DeepSeek. Install first with npm install -g @anthropic-ai/claude-code (Node.js 18+ required; Windows also needs Git for Windows).

The official config gives heavy reasoning to Pro and cheap high-frequency work to Flash. Flash carries the Haiku tier and sub-agents; Pro carries the main task. The [1m] suffix selects the 1M-context variant. On Windows, write each export with the $env: form. Verify the install with claude --version.

Web Search works natively: Claude Code calls DeepSeek's Web Search tool when the model decides it is needed, and summarizing the results adds token cost. The awesome-deepseek-agent repo adds one more variable, CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1^[8], and writes the Haiku tier as deepseek-v4-flash[1m].

• Community reports put Claude Code on DeepSeek at roughly $7/month instead of $200/month using just two environment variables (MG-Cafe, live-tested, not official numbers).
• The same mapping lets Claude Desktop connect: enable developer mode, change the base URL and API key, and the claude-* names bypass the app's model-name restrictions.

npm install -g @anthropic-ai/claude-code

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=<your DeepSeek API Key>
export ANTHROPIC_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash
export CLAUDE_CODE_EFFORT_LEVEL=max
cd /path/to/my-project
claude

The 0731 build of DeepSeek V4 Flash natively supports the OpenAI Responses API^[2], which is the protocol Codex speaks. That ends the proxy era: one official script wires up Codex, and it is re-runnable to switch models or restore the original config.

Codex CLI, the ChatGPT desktop app, and the VS Code Codex extension share the same ~/.codex/ configuration. Today only deepseek-v4-flash integrates with Codex; DeepSeek says deepseek-v4-pro support arrives in early August 2026.

The Responses endpoint has no previous_response_id, so multi-turn sessions resend the full context, and OpenAI built-in tools like file_search and code_interpreter are not supported.

• Backs up ~/.codex/config.toml to ~/.codex/backup-deepseek/.
• Writes ~/.codex/models.json declaring the model card: context window 1,048,576, reasoning levels low/high/max, and the apply_patch tool type.
• Adds a [model_providers.deepseek] section to config.toml and keeps your existing MCP and project-trust settings.
• Validates syntax before writing and aborts without touching any file on failure.
• Prompts for your sk- API key on first run.

# macOS / Linux — run once after installing Codex CLI
bash <(curl -fsSL https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.sh)

# Windows PowerShell
irm https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.ps1 | iex

The Anthropic endpoint maps claude-* model names onto DeepSeek models automatically. In most Claude Code sessions you never type a DeepSeek slug at all.

That is why the official env config sets ANTHROPIC_DEFAULT_HAIKU_MODEL to deepseek-v4-flash: Claude Code still routes its internal haiku-tier calls, and DeepSeek serves them with Flash.

The mapping also explains the cost math. Flash carries the haiku tier and sub-agents, so only the main task runs on the more expensive pro model.

Flash and Pro are two tiers of one model family. Flash is the fast, cheap, high-frequency tier; Pro is the flagship for complex reasoning. Pick by task, not by habit.

DeepSeek's own benchmarks, run on its DeepSeek Harness in minimal mode with max effort, put Flash-0731 at 82.7 on Terminal-Bench 2.1 and 54.4 on DeepSWE. The changelog says all nine published agent benchmarks beat V4-Pro-Preview. Treat these as vendor-reported until third parties reproduce them. On the independent Artificial Analysis Intelligence Index v4.1, V4-Flash-0731 scores 50, ranking #2 of 162 models in its price tier.

A typical Cursor Composer run is around 20K input and 3K output tokens: about $0.013 on Pro and $0.003 on Flash. A full day at roughly 150 rounds is about $2 on Pro and under $0.50 on Flash. Community estimates put a $5 top-up at about a week of heavy Flash usage.

The most time-consuming Cursor issue is the reasoning_content 400 on long agent runs. DeepSeek thinking mode returns both content and reasoning_content in every streaming chunk. Cursor's Chat panel renders both; Composer renders only content and throws HTTP 400 on long tool chains.

The 400 happens because DeepSeek thinking mode requires the next request to replay the full reasoning_content chain. When Cursor packages the next round, it strips that field, so DeepSeek rejects the request. Three fixes exist, from cheapest to most complete.

• Use Chat to read the reasoning and Composer to change files. Zero cost, and the recommended default.
• Turn off thinking mode in Composer or switch to Flash. You lose the reasoning gain, but the agent stays stable.
• Run the community proxy yxlao/deepseek-cursor-proxy^[7]. Point Cursor at http://localhost:<port>; the proxy caches the chain-of-thought per session and replays it on the next round. The full Composer agent loop works, and thinking tokens display as collapsed Markdown.