DeepSeek V4 Pro Review: Agent Leap Is Real, So Is the New Bill

V4 Pro's DeepSWE score jumped from 12.8 to 62.7 and it ties Claude Fable 5 on Terminal Bench — but the new peak/off-peak pricing raises cache-hit costs 12x. Full review with the real cost math.

🚀 A Silent Release With Loud Numbers

DeepSeek shipped the official V4-Pro-0813 at 3 a.m. on August 13, 2026 — no launch blog, just an API pricing-page update and a quiet OpenRouter refresh. The architecture did not change one bit: still a 1.6-trillion-parameter MoE with about 49B activated parameters. Everything that improved came from targeted post-training, and the gains landed exactly where it matters most right now: agents.

The before/after is dramatic. On DeepSWE, a software-engineering agent benchmark, V4 Pro jumped from 12.8 to 62.7 — a 390% improvement. NL2Repo rose from 38.5 to 61.5, and DSBench-Hard hit 67.2. On Terminal Bench it scored 87.9 versus Claude Fable 5's 88.0 — a statistical tie with the model many consider the coding benchmark to beat. Against Fable 5 across nine comparable benchmarks, DeepSeek still trails by about 5.3% on average, but it wins CyberGym and AutomationBench outright.

The practical takeaway: 62.7 on DeepSWE crosses the line from "assisted coding" to "autonomous coding". This is now a model that can plan, edit across files, and self-correct on real repositories — not just autocomplete with extra steps.

📏 1M Context: The Spec Sheet Only Two Models Can Match

V4 Pro accepts up to 1 million input tokens and can generate up to 393,216 output tokens in one go. That is 8x the 128K window of V3.2 and 4x Kimi K2.6's 256K; among open-weight flagships, only MiniMax M1 plays in the same tier. In practice, you can drop an entire front-end codebase in for global analysis and still get a complete scaffold back without the output truncating halfway.

💰 The Part Everyone Missed: The New Bill

At launch, V4 Pro cost ¥3 per million input tokens (cache miss), ¥0.025 cache-hit input, ¥6 output — with concurrency capped at 500 (versus 2,500 for V4 Flash). Four days later, on August 17 at 00:00 Beijing time, the new pricing took effect: peak hours (9:00–12:00, 14:00–18:00) now cost ¥9 input / ¥27 output / ¥0.30 cache-hit; the other 17 hours are half that (¥4.5 / ¥13.5 / ¥0.15).

The percentages sting: peak output +350%, cache-hit input +1,100% — twelve times. The group hit hardest is not casual users but coding agents that repeatedly ship the same repository, specs, and conversation history. A representative workload — 10M input plus 2M output tokens, no cache hits — now costs about ¥72 off-peak or ¥144 peak. The same workload on Claude Fable 5 runs roughly ¥1,440.

So the verdict on price is nuanced. At ¥27 peak output, V4 Pro matches Gemini 3.7 Flash's converted rate and sits ¥3 below Doubao Seed 2.1 Pro — while remaining about one-eighth of GPT-5.6 Sol and one-seventh of Claude Opus 4.8. DeepSeek is still cheap; it is just no longer "cheap enough that you skip the spreadsheet".

🔌 Ecosystem: Dual Protocol Changes the Switching Cost

The most strategically interesting change is invisible in benchmarks: the API now speaks both the Responses API and the Anthropic API protocol, making V4 Pro a drop-in for tools built around Codex and Claude Code. Add the debut of the DeepSeek Harness minimal mode for agent orchestration, and switching your agent stack to V4 Pro becomes an afternoon of config rather than a rewrite. The demand is clearly there: on OpenRouter's weekly token leaderboard, DeepSeek V4 Flash (0731) ranked #1 globally at 8.83 trillion tokens, up 570% week-over-week.

⚠️ Before You Migrate Everything

Two cautions. First, reliability: DeepSeek suffered outages on May 8, 21, 24, and 28, caused by user traffic growing 66.7% against compute growth of just 8.3% — the peak/off-peak pricing is explicitly a demand-shaping response. Design for retries and off-peak batch jobs. Second, route sensibly: simple tasks belong on V4 Flash, complex ones on Pro, and anything latency-insensitive should be scheduled into the 17 off-peak hours at half price.

Verdict: a 4.5 out of 5. The agent leap is real and the 1M context is genuinely useful; the pricing restructure means engineering teams should re-run their cost models instead of assuming "DeepSeek is the cheap option". For agent and coding workloads with off-peak scheduling, it remains one of the best value-per-yuan flagships available.