DeepSeek V4 Flash Gains Native Codex Support: Why Users Call It “Pure”

This isn’t just “OpenAI‑compatible”

Previously, integrating DeepSeek with Codex required a conversion layer because DeepSeek’s API offered only OpenAI Chat Completions and Anthropic‑compatible formats, while Codex uses the Responses API. The extra translation step introduced maintenance overhead and potential breakage when either side changed fields.

V4 Flash adds native support for the Responses API. The base URL remains https://api.deepseek.com, so a Codex request can be sent directly to DeepSeek without any intermediate wrapper. A one‑click configuration script can also populate ~/.codex/config.toml and ~/.codex/models.json, allowing Codex CLI, the ChatGPT desktop client, and the VS Code extension to share the same settings.

DeepSeek still backs Claude Code

DeepSeek continues to provide Anthropic‑compatible endpoints for Claude Code, with environment variables such as ANTHROPIC_BASE_URL and API‑key mappings. V4 Pro and V4 Flash map to different Claude model tiers, even offering configuration examples for sub‑agents.

However, that route depends on Claude Code’s compatibility boundaries; any change in Claude’s client rules would require adjustments. By contrast, DeepSeek now implements the Responses API itself, exposing capabilities like apply_patch, function calls, web search, and adjustable inference strength, aligning directly with Codex’s workflow.

Flash focuses on Agent capabilities

V4‑Flash‑0731 keeps the same model architecture and scale as the preview release; the main changes come from post‑training, especially improving Agent abilities.

Benchmark results on Terminal Bench 2.1 show the following scores:

V4 Flash: 82.7 (up from 61.8 in the preview)

V4 Pro Preview: 72.1

GLM‑5.2: 81.0

Opus 4.8 (official table): higher than all above

Other internal tests show NL2Repo rising from 39.4 to 54.2, DeepSWE from 7.3 to 54.4, DSBench‑FullStack at 68.7, and DSBench‑Hard at 59.6. These internal metrics illustrate the version’s progress but are not meant for cross‑vendor comparison.

The Responses API now covers Codex’s common function calls such as web_search and apply_patch. Stable operation in a terminal depends more on the model’s ability to read the environment, manipulate tools, and edit files than on raw code‑generation skill—precisely what the latest post‑training targets.

Pricing is “pure” according to the community

Netizens praise DeepSeek’s straightforward pay‑per‑token model. At the current official rates, V4 Flash costs $0.0028 per million tokens for cache‑hit input, $0.14 for cache‑miss input, and $0.28 per million tokens for output.

This pricing is attractive for Codex scenarios where tools repeatedly read code, tool results, and context, leading to many API calls per task. The cost model remains transparent: no subscription tiers, no hidden quotas—just real‑time token consumption.

However, DeepSeek has announced upcoming peak‑off‑peak pricing (double the normal rate during 9:00‑12:00 and 14:00‑18:00), so the “pure” simplicity may evolve.

Limitations and target users

Only the deepseek-v4-flash model currently supports Codex; V4 Pro cannot use the Responses API yet (expected support in early August 2026). The API still lacks image and file inputs, previous_response_id, background tasks, Code Interpreter, Computer Use, and MCP tools. Unsupported parameters are ignored rather than causing errors.

V4 Flash is in public testing with a context window of 1 million tokens and a maximum output of 384 k tokens. Stability for very long contexts, complex toolchains, and large repositories still needs real‑world validation.

Thus, the model best fits two groups: developers already using Codex who want lower per‑call costs, and teams that need to batch‑process text and code without requiring multimodal inputs or the missing built‑in tools. Projects heavily reliant on images, files, or long‑running background tasks should consider alternative models.

Overall, DeepSeek is handing the choice of front‑end tools back to developers: they can keep their existing Codex workflow while swapping the underlying model, paying only for actual usage.