DeepSeek V4-Pro is officially released, now versioned DeepSeek-V4-Pro-0813 — with pricing, specs, and the API call name all unchanged. It's a "preview to official" graduation: both models in the V4 lineup are now fully official (source).
What Actually Changed
V4-Pro graduated from preview, and the official pricing page now lists DeepSeek-V4-Pro-0813. The July 31 changelog entry already said the V4-Pro official release "will follow soon" (source). The version bump on the pricing page means it's now live. That's the whole story: the version number changed, nothing else did.
| Item | Before | Now (official) |
|---|---|---|
| Model version | V4-Pro preview | DeepSeek-V4-Pro-0813 |
| API call name | deepseek-v4-pro | deepseek-v4-pro |
| Input (cache miss) | $0.435 / MTok | $0.435 / MTok |
| Output | $0.87 / MTok | $0.87 / MTok |
| Input (cache hit) | $0.003625 / MTok | $0.003625 / MTok |
| Concurrency | 500 | 500 |
Enhanced Agent Capability + Codex Integration
DeepSeek confirmed the V4-Pro official release enhances Agent capability and adds Responses API and Codex integration. The official announcement from DeepSeek's user community reads: "V4-Pro official release is now live on the API, model name unchanged, with enhanced Agent capability and Responses API + Codex integration — welcome to test and provide feedback." Responses API support is already visible on the pricing page (source), and Codex integration is now in the official docs — both deepseek-v4-pro and deepseek-v4-flash work directly in Codex CLI, the ChatGPT desktop app, and the VS Code Codex extension (source).
As for how much the Agent capability actually improved, DeepSeek hasn't published benchmark numbers yet, and the changelog hasn't caught up to this Pro entry. So the confirmed story is "capability up + Codex works" — the specific scores still need official data.
Pricing and Specs: Nothing Moved
This update involves no price change — pricing and specs are identical to before. Here's the current config for both models on the official pricing page (source):
| Item | V4-Flash | V4-Pro |
|---|---|---|
| Input (cache hit) | $0.0028 / MTok | $0.003625 / MTok |
| Input (cache miss) | $0.14 / MTok | $0.435 / MTok |
| Output | $0.28 / MTok | $0.87 / MTok |
| Concurrency | 2500 | 500 |
| Context | 1M | 1M |
| Max output | 384K | 384K |
Features are identical across both models too: thinking/non-thinking modes, JSON Output, Tool Calls, Responses API, and Anthropic API all supported; FIM completion is non-thinking mode only.
How to Choose Between V4-Flash and V4-Pro
Daily work goes to Flash, complex reasoning goes to Pro — that advice hasn't changed. Pro costs 3x more and has 5x lower concurrency, but what you're buying is stronger deep-reasoning capability. If your workload is complex refactors, long multi-step agent chains, or knowledge-intensive workflows where every step depends on the last, the 3x premium is worth it. Otherwise — daily coding, batch jobs, light agents — Flash is plenty.
Things to Watch
- No official benchmarks yet. The changelog still stops at the July 31 Flash entry. Pro's official technical report and agent benchmarks aren't out, so don't compare against the old preview scores.
- The price-hike warning is still live. The official pricing page explicitly says DeepSeek "plan[s] to raise the overall pricing ... with a significant increase expected." Get your caching strategy and usage estimates sorted before it hits.
- No code changes needed. Existing deepseek-v4-pro calls route to the new version automatically — no code changes, no key rotation.
- Prefer a fixed monthly fee over pay-as-you-go? Check out third-party plans from Volcengine and CtCloud, which aren't subject to the official concurrency cap or price hike.