BREAKING (July 31, 2026): DeepSeek V4 Flash Official API Goes Public Beta! Agent Performance Surpasses V4 Pro Preview — Native Support for Claude Code, Codex, OpenCode, Hermes, and 6 More Tools

It's finally here. DeepSeek's official V4 Flash API has been released in public beta.

On July 31, 2026, DeepSeek announced it on X. The biggest shock: "V4 Flash's agent performance has far surpassed V4 Pro Preview." In other words, the free-to-ultra-cheap Flash model has overtaken the top-tier model.

It also has native support for 6 major agent tools: Claude Code, Codex, OpenCode, OpenClaw, Hermes, and Reasonix. It supports the Responses API format and is fully optimized for Codex.

This article explains pricing, performance, supported agents, and setup steps for beginners, based on real data from DeepSeek's official X announcement and the official API documentation.

Here's what you'll learn:

• How much V4 Flash surpasses V4 Pro Preview
• Pricing per 1M tokens (from $0.0028 on cache hits)
• How to configure it in Claude Code, Codex, OpenCode, Hermes, and 6 other tools
• OpenAI-compatible and Anthropic-compatible base_url and model names
• Peak/off-peak pricing cautions
• Feature list including Responses API and Thinking Mode

The Core of the Announcement (Full X Post)

DeepSeek official (@deepseek_ai) post:

🚀 DeepSeek-V4-Flash Official API is now LIVE in public beta!

🔷 We've massively upgraded its Agent capabilities — benchmark scores are now far surpassing the V4-Pro-Preview.

🔷 The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex!

Translation:

• V4 Flash's agent performance has dramatically improved, with benchmark scores far surpassing V4 Pro Preview
• Native support for the Responses API format
• Fully optimized for Codex

Massive response: 5.75M views and 1,348 replies. Unsloth AI also reacted.

Pricing (August 2026, per official API docs)

V4 Flash's cost performance is overwhelming. On cache hits, input is $0.0028 (about ¥0.4)/1M tokens. Compared to V4 Pro, input is about 1/3 and output about 1/3.

⚠️ Peak/off-peak pricing is coming soon. Peak hours (Beijing time 9:00-12:00, 14:00-18:00, UTC+8) will be 2x. In Japan time: 10:00-13:00, 15:00-19:00.

Spec Comparison (V4 Flash vs V4 Pro)

Key point: Responses API is Flash-only (V4 Pro support expected early August). So if you want to use Codex now, V4 Flash is the only choice.

Supported Agents (6 Tools, per official API docs)

All supported tools listed in the "Agent Integrations" section of the official API docs:

Setup (Done in 3 Minutes)

Common Setup

# Get an API key
# Register at https://platform.deepseek.com/ and issue a key

# Set the environment variable
export DEEPSEEK_API_KEY="sk-xxxxxxxxxxxxxxxx"

① Using with Claude Code

# Configure as an Anthropic-compatible API
claude config set api_key $DEEPSEEK_API_KEY
claude config set api_base https://api.deepseek.com/anthropic
claude config set model deepseek-v4-flash

② Using with Codex (native Responses API support)

# Codex directly supports the DeepSeek Responses API
codex config set provider deepseek
codex config set model deepseek-v4-flash

③ Using with OpenCode / OpenClaw / Hermes / Reasonix

# Configure as an OpenAI-compatible API
export OPENAI_API_KEY=$DEEPSEEK_API_KEY
export OPENAI_BASE_URL="https://api.deepseek.com"

# Specify the model name
# OpenCode: opencode --model deepseek-v4-flash
# Hermes: hermes config set provider_model deepseek-v4-flash

④ Calling the API directly (curl)

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": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello!"}
    ],
    "thinking": {"type": "enabled"},
    "reasoning_effort": "high",
    "stream": false
  }'

Pricing Comparison with Other Models (per 1M output tokens)

V4 Flash is about 1/50th of GPT-5.6 Sol and about 1/270th of Claude Opus 4.8. This price disruption is its biggest weapon.

Cautions (from official docs)

⚠️ Peak/off-peak pricing (coming soon)

• Peak hours (Japan time 10:00-13:00, 15:00-19:00) are 2x
• Using the API late at night/early morning is roughly half price
• Actual launch date TBD ("to be announced soon")

⚠️ Responses API is Flash-only

• V4 Pro Responses API support is "expected early August"
• If you use Codex, V4 Flash is the only choice (for now)

⚠️ Thinking Mode is ON by default

• Thinking tokens are included in billing
• Turn it off with thinking: {type: "disabled"} if unwanted

⚠️ Per-account concurrency limits

• Flash: 2500, Pro: 500

FAQ

Q1. Can I use it for free?

Not completely free, but a $5 top-up lasts months. On cache hits, input is $0.0028 per 1M tokens.

Q2. Does it really work with Claude Code?

Yes. Officially supported as an Anthropic-compatible API.

Q3. V4 Pro or V4 Flash — which should I choose?

Flash for cost performance. Pro for maximum performance. But Responses API (for Codex) is Flash-only.

Q4. Does it support Japanese?

Yes. Multilingual support.

Q5. Where do I get an API key?

Register at https://platform.deepseek.com/ — issued instantly.

Q6. Can I use it in other countries?

Available globally (excluding some regions).

Q7. What is Thinking Mode?

The model "thinks" internally before answering. More accurate, but output tokens increase.

Q8. What model name should I specify for V4 Flash?

deepseek-v4-flash (auto-routes to the latest DeepSeek-V4-Flash-0731).

Summary — August 2026: The "Standard Engine" for AI Agents Is Decided

Conclusion: The DeepSeek V4 Flash official API is the standard engine for AI agents in the second half of 2026 — price disruption, agent performance, and tool support all in one.

• Price: $0.28 per 1M output (about 1/50th of GPT-5.6 Sol)
• Performance: agent benchmarks surpassing V4 Pro Preview
• Support: Claude Code, Codex, OpenCode, OpenClaw, Hermes, Reasonix — 6 tools
• Setup: 3 minutes with the OpenAI-compatible API

Start right now:

# Get an API key and
curl https://api.deepseek.com/chat/completions \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"Hello"}]}'

"Flash surpassed Pro." That's the front line of AI in August 2026.

Try the Setup from This Article Right Now

If you want to actually touch the DeepSeek V4 Flash official API, OpenCode Go (from $10/mo) is the fastest way. All models including DeepSeek V4 Flash are unlimited, and the setup steps introduced in this article work as-is. No need to top up API keys separately.

💡 Signing up through the referral code JQR0M51199 gives $5 of credit to both you and me. Start with the feeling of trying it for a month.

Information in this article is as of August 1, 2026. Check the DeepSeek official API docs for the latest pricing and specs.