DeepSeek made V4 Flash official on July 31, and the buried headline is this: the smaller model now beats the larger one. V4-Flash-0731 outperforms V4-Pro-Preview across all nine published agentic benchmarks — according to DeepSeek’s own release notes. Terminal Bench 2.1 sits at 82.7 for Flash versus Pro’s 72.1. DeepSWE lands at 54.4 versus 12.8. Cybergym: 76.7 versus 52.7. The smaller model wins every category, at a third of the output cost. If you’re running coding agents and haven’t switched yet, here’s your migration guide.
What Changed From Preview to GA #
Almost nothing architectural changed — and that’s the interesting part. DeepSeek re-ran only the post-training phase. Same 284B MoE architecture, same 13B active parameters per token, same 1M context window. The model string stays deepseek-v4-flash
, so existing integrations require no code changes. What did land in this build: native Responses API support, which makes it directly compatible with Codex-based integrations out of the box.
One housekeeping note before setup: if you’re still calling deepseek-chat
or deepseek-reasoner
, those aliases retired on July 24. Switch to deepseek-v4-flash
now or your calls are failing silently.
The Benchmark Numbers Worth Knowing #
| Benchmark | Flash-0731 | Flash Preview | Pro Preview |
|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 61.8 | 72.1 |
| DeepSWE | 54.4 | 7.3 | 12.8 |
| Cybergym | 76.7 | 38.7 | 52.7 |
| NL2Repo | 54.2 | 39.4 | 38.5 |
| Toolathlon (verified) | 70.3 | 49.7 | 55.9 |
On DeepSWE — a real-world software engineering benchmark — Flash is more than four times better than its preview version and four times better than Pro Preview. Independent verification from Artificial Analysis puts Flash at #2 of 162 models measured. One developer in the HackerNews discussion put it plainly: “It would absolutely have been a frontier model last December.”
What It Costs — Honestly #
| Token Type | V4-Flash | V4-Pro |
|---|---|---|
| Input (cache miss) | $0.14/M | $0.435/M |
| Input (cache hit) | $0.0028/M | $0.003625/M |
| Output | $0.28/M | $0.87/M |
Output tokens cost 3x less than Pro. Cache hits knock 98% off input costs — critical for agentic loops where your system prompt repeats on every tool call. In practice, a full day of agent-heavy work runs under $0.50 on Flash versus roughly $2 on Pro.
There’s a real caveat here: Flash generates approximately three times the median token volume on agentic tasks due to its internal reasoning chain. Effective cost per completed task runs higher than raw rates suggest — but it’s still cheaper than Pro in every real-world scenario. Also worth noting: DeepSeek has announced a 2x peak-hour surcharge during Beijing business hours, effective date still TBD.
Setting Up in Claude Code #
DeepSeek provides an Anthropic-compatible endpoint, so setup is a handful of environment variables. The official integration docs cover it, but the smart mapping is to assign Flash to lighter slots (Haiku, subagents) and Pro to the main model role:
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=your_deepseek_api_key
export ANTHROPIC_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash
export CLAUDE_CODE_EFFORT_LEVEL=max
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=786432
This maps Flash to the high-frequency sub-tasks and reserves Pro for the primary reasoning pass. Given that Flash outperforms Pro on every agentic benchmark, you could map everything to Flash — but the hybrid approach is the sensible starting point while Pro’s GA is still pending.
Setting Up in Cursor #
Open Settings (Cmd+, on Mac, Ctrl+, on Windows), go to Models, and toggle “Override OpenAI Base URL.” Enter https://api.deepseek.com
— no trailing /v1
. Paste your DeepSeek API key into the OpenAI key field, click “+ Add model,” type deepseek-v4-flash
, and hit Verify. Switch to it in the chat model picker before starting a session.
There’s one important caveat here: Cursor’s Composer panel has a known issue with extended tool-call sequences. Cursor only forwards the content
field to the model — not reasoning_content
— which causes failures when DeepSeek’s thinking mode needs to replay its chain across multiple requests. Use the Chat panel for complex multi-step tasks or disable thinking mode for Composer sessions. Background Agents also don’t support custom models as of this writing.
Get Your API Key #
If you don’t have one, start at the DeepSeek platform. Minimum wallet funding is $5 — that covers weeks of normal agent usage at Flash pricing.
The Takeaway #
V4-Flash is the agent model of summer 2026. A frontier-adjacent model at sub-commodity pricing, MIT-licensed, open weights available for self-hosting on 4x A100s. V4-Pro still hasn’t hit GA — when it does, the picture may shift. For now, Flash is the default choice for coding agents. The benchmarks say so, and so does your wallet.