DeepSeek V4 Flash Lands with Native Codex Support and a Config Fix for CLI 0.137.0
DeepSeek V4 Flash is the first model from the company to ship with a turnkey Codex integration script, lowering the friction for developers who want a non-OpenAI provider for agent workflows. The CLI 0.137.0 compatibility snag is a concrete, fixable breakage that anyone following the official setup will hit immediately.
DeepSeek released the V4 Flash model into public beta on July 31, 2026, shipping an official PowerShell and bash script that rewires Codex to use it as the default provider. The model keeps the `deepseek-v4-flash` name and targets agent workloads, with vendor-reported scores of 82.7 on Terminal Bench 2.1 and 70.3 on Toolathlon verified. The setup script backs up existing config, writes a new `models.json`, and points the provider at DeepSeek's Responses API endpoint.
Codex CLI 0.137.0 rejects the generated config because it sets reasoning effort to `max`, a value the parser does not recognize. A targeted string replacement to `xhigh` clears the error, and `codex doctor --json` then confirms the model, provider, and wire API are correctly loaded. The fix is specific to that CLI version; later releases may not need it.
After configuration, the desktop client shows a `custom` badge but that alone does not prove connectivity. A real task with a valid API key is still required to confirm latency, output quality, and token consumption. The article also flags that agent runs with Codex generate long contexts and many tool calls, so list prices can mislead; comparing actual token burn across channels gives a truer cost picture.
The `max` reasoning-effort mismatch is a version-skew problem between DeepSeek's script output and Codex's parser, not an API or model bug, and it will likely vanish once Codex updates its accepted values.
Vendor benchmarks like Terminal Bench 2.1 and Toolathlon are useful for rough positioning but say nothing about latency, token economics, or failure modes under real agent workloads.
Comparing actual token consumption across different API channels for the same task is a more reliable cost-control tactic than trusting any single provider's list price.