Overnight DeepSeek V4 Pro Test Reveals Disappointing Performance – Not Fit for Codex

DeepSeek V4 Pro was officially released with unchanged pricing and immediate API access.

The model card shows capabilities roughly on par with kimi‑k3 and Fable 5, yet priced at a floor level.

Integrating V4 Pro into Codex, the author processed 41 million tokens and concluded that it is not suitable as the main model.

1. Silent, Immersive Operation

Unlike the chatty Gemini Flash, V4 Pro runs silently for twelve to thirteen minutes, providing almost no intermediate status. In the test, 84.5% of the output tokens were devoted to thinking, with no process commentary.

2. Extreme Context Craving and Copy‑Paste Behavior

The model aggressively reuses traces from previous sessions, preferring to copy existing solutions rather than generate from scratch. When given a 3D FPS multiplayer game requirement, it returned an existing implementation from the author’s prior Flash test and simply launched it. For a trading‑strategy backtest, it claimed to have rerun the local engine, reusing earlier results.

3. Frontend Generation Below Expectations

The quality of generated frontend code and rendered content is noticeably poorer than anticipated. Attempts to visualize pressure‑test data resulted in crude graphs, and other visual outputs (e.g., the “pelican riding a bike” demo) were consistently subpar compared to peer models.

4. Writing Ability Below Expectations

When fed the same narrative prompt to four models, the ranking was Gemini 3.1 Pro > GPT‑5.6 > DeepSeek V4‑Pro > Gemini Flash. DeepSeek produced occasional good lines, but overall the text was thin, contained many irrelevant environmental descriptions, and lagged in instruction following.

5. Strong Fundamentals Yet Higher Latency

Basic performance tests show excellent stability: 100 concurrent requests with 0 failures, 100% success rate, total throughput 52.3 K tok/s, output speed 4.6 K tok/s, RPS 10.4, and cache‑hit rate 97.5%. The author’s billing recorded 41.39 M tokens, 39.72 M of which hit cache, costing $8.5. However, latency is higher than Flash, with median (p50) 4.9 s and 99th‑percentile (p99) 12.3 s. Cost‑wise, Flash can process about 100 M tokens per hour for $8, whereas V4 Pro processes less than half that volume for the same price.

Conclusion

V4‑Pro is a capable model but behaves like a quiet, copy‑prone teammate that underperforms in frontend generation and creative writing. While its throughput and cost remain attractive, the higher latency and tendency to reuse prior results make it a risky primary choice for Codex.