DeepSeek V4‑Flash Official Release: Direct Codex Integration and Benchmark Gains

Official Release and Version Change

On July 31 DeepSeek released the official DeepSeek‑V4‑Flash API. The model name remains deepseek‑v4‑flash and the call signature is unchanged, but the backend switched from the April preview to the 0731 version. Regression tests must account for this version change.

Scope of the Update

The change affects only the V4‑Flash API. V4‑Pro, APP and WEB endpoints are unchanged. Model architecture and parameter count are identical to the preview; the improvement comes from a full re‑training focused on agent tasks.

Older model names deepseek-chat and deepseek-reasoner were deprecated on July 24. Projects that hard‑code these names should update them now.

Benchmark Results

DeepSeek reports nine benchmark scores obtained with the proprietary DeepSeek Harness minimal mode (parameters: top_p=0.95, temperature=1.0). The harness is not publicly released, so scores may differ with other frameworks.

Terminal Bench 2.1: 82.7

NL2Repo: 54.2

Cybergym: 76.7

DeepSWE: 54.4

Toolathlon verified: 70.3

Agent Last Exam: 25.2

Automation Bench (Public): 25.1

DSBench‑FullStack: 68.7

DSBench‑Hard: 59.6

Two annotations clarify the numbers:

The benchmark suite runs on DeepSeek’s internal DeepSeek Harness minimal mode with top_p=0.95 and temperature=1.0. Scores may not transfer directly to Claude Code or Codex pipelines.

DSBench‑FullStack and DSBench‑Hard are internal full‑stack and hard coding‑agent test sets; they are not public leaderboards, so cross‑model comparison is not meaningful.

Terminal Bench 2.1 (82.7) is highlighted as the most relevant to real‑world coding agents because it measures continuous terminal operations and adaptive step‑by‑step reasoning.

Codex Integration

V4‑Flash now natively supports the Responses API and includes a Codex‑specific configuration.

One‑click script (creates a backup of ~/.codex):

# Bash one‑liner (creates backup)
bash <(curl -fsSL https://cdn.deepseek.com/api-docs/codex-deepseek-setup.sh)

PowerShell equivalent:

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

Manual configuration (two steps):

# Create ~/.codex/models.json with model metadata (JSON omitted for brevity)
# Edit ~/.codex/config.toml
model = "deepseek-v4-flash"
model_provider = "deepseek"
preferred_auth_method = "apikey"
forced_login_method = "api"
model_reasoning_effort = "high"
model_catalog_json = "~/.codex/models.json"

[model_providers.deepseek]
name = "deepseek"
base_url = "https://api.deepseek.com/"
wire_api = "responses"
experimental_bearer_token = "<your API Key>"

Key fields: wire_api must be responses to enable the Responses API. model_reasoning_effort selects the reasoning level (higher = slower but higher quality). experimental_bearer_token accepts the API key directly, bypassing the login flow.

Only deepseek-v4-flash supports Codex today; support for deepseek-v4-pro is planned for August 2026.

Claude Code Integration

Anthropic‑compatible environment variables point to DeepSeek endpoints. Default heavy‑weight models (Sonnet, Opus) target

deepseek-v4-pro</>, while the high‑effort Claude Code sub‑agent uses <code>deepseek-v4-flash

for cost‑effective performance.

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=<your 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

Web‑search calls incur extra token fees because the retrieved content must be summarized.

Responses API Details

Only deepseek-v4-flash currently supports the Responses API; deepseek-v4-pro will follow in August. Base URL: https://api.deepseek.com. Example using the OpenAI SDK:

from openai import OpenAI
client = OpenAI(api_key="<your API Key>", base_url="https://api.deepseek.com")
resp = client.responses.create(
    model="deepseek-v4-flash",
    instructions="You are a helpful assistant.",
    input="Hi, how are you?",
)
print(resp.output_text)

Supported parameters include tools, tool_choice, reasoning, text, temperature, top_p, max_output_tokens, and stream. The following are silently ignored:

previous_response_id
conversation
store
background
metadata
prompt_cache_key
service_tier

The API is stateless; the full conversation must be supplied in input. Parallel tool calls are always enabled and cannot be disabled. Streaming returns semantic SSE events ( response.created, response.output_text.delta, response.completed) instead of the [DONE] marker used by ChatCompletions.

Pricing and Peak‑Time Considerations

Both models offer 1 M context and a 384 K maximum output. Costs (CNY per million tokens):

deepseek‑v4‑flash : input (cache hit) 0.02, input (miss) 1, output 2

deepseek‑v4‑pro : input (cache hit) 0.025, input (miss) 3, output 6

Cache‑hit input cost is critical for coding agents that repeatedly feed the same files; a high cache‑hit rate makes input cost negligible.

Peak hours (Beijing time 09:00‑12:00 and 14:00‑18:00) double the price. Running large‑scale evaluations overnight can halve expenses.

Key Takeaways

The update targets agent capability rather than raw knowledge.

Native Responses API support and Codex adaptation remove an extra translation layer.

Benchmark scores rely on an unreleased harness; replication may vary.

Codex and Responses API currently work only with deepseek-v4-flash; deepseek-v4-pro support is slated for August 2026.

V4‑Pro’s official release date is not announced.

Recommended migration path: test deepseek-v4-flash on a small subset of real tasks via Codex or as a Claude Code sub‑agent before a full rollout.

Reference Links

Update log: https://api-docs.deepseek.com/zh-cn/updates

Codex integration: https://api-docs.deepseek.com/zh-cn/quick_start/agent_integrations/codex

Claude Code integration: https://api-docs.deepseek.com/zh-cn/quick_start/agent_integrations/claude_code

Responses API guide: https://api-docs.deepseek.com/zh-cn/guides/responses_api

Model pricing: https://api-docs.deepseek.com/zh-cn/quick_start/pricing