DeepSeek V4 Pro Review: The GA Build Under a Microscope

DeepSeek V4 Pro 0813 is the strongest open-weight model for agentic coding as of mid-August 2026, and the GA build's improvement over the April preview is one of the largest single-version jumps on record. It is not, however, the right model for every job — independent tests flag over-thinking on simple tasks, and the new peak-hour pricing raises the cost of always-on workloads^[4][6].

This review draws on the official model card, MindStudio's independent 8-task evaluation (August 13), community threads, and DeepSeek's own pricing and GA announcements. Where numbers disagree between vendor and independent sources, both are shown^[1][4][6].

One sentence summary: buy it for agentic coding and hard reasoning; route everyday chat and simple tasks to V4 Flash.

• Agentic coding: best-in-class among open models (TB2.1 87.9, Cybergym 83.3 first)
• Independent score: 76.25% on MindStudio's suite (preview: 24.8%)
• Pricing: $1.98/$3.96 per 1M output tokens (off-peak/peak) from Aug 16
• Best used with: Codex via Responses API, harness-driven agent loops

The GA build is a coding and agent powerhouse. On DeepSeek's official tables it takes first place on Cybergym (83.3) and AutomationBench (31.8), and its Terminal Bench 2.1 score of 87.9 sits just below Opus-4.8's 88.3^[4].

Independent testing confirms the direction: MindStudio's suite put the 0813 build at 76.25%, roughly tied with Muse Spark 1.2 and within reach of Kimi K3 and Opus 5 — remarkable for a fully open-weight model^[6].

The full benchmark page has every number with the harness caveat explained.

• Agentic coding: elite terminal-bench and tool-use scores, produced with the DeepSeek Harness in minimal mode at max effort
• Software engineering: DeepSWE 62.7 — the biggest single gain over the preview (+49.9)
• Front-end generation and task planning (flagged as strengths in independent testing)^[6]
• Long-context: native 1M-token context with a 384K max output
• Responses API: first-class Codex integration via one-click setup^[5]

The same independent test that praised V4 Pro's front-end work flagged two behavioral quirks: it over-thinks simple problems, and it rewrites code more than the task needs^[6].

These weaknesses are manageable with configuration. The reasoning effort guide shows how `low` effort cuts the over-thinking problem, and the context caching guide explains how to keep the bill flat for agent loops.

Community threads echo the same pattern — 'sometimes the little brother V4 Flash is better for daily work' is a recurring sentiment^[6][9].

• Over-thinking: simple requests get long reasoning chains, adding latency and output-token cost
• Over-rewriting: when asked for a small change, it often regenerates more than necessary
• Peak-hour pricing: the new $3.96/M output peak rate makes idle-until-midnight bots expensive
• Self-hosting: ~862GB VRAM estimate makes local deployment a data-center project
• Concurrency: 500 concurrent requests — well below Flash's 2,500

Value math changed on August 16, when peak/off-peak pricing replaced the flat rate. Every V4 Pro line is now more expensive than the old $0.435/$0.87 flat pricing^[2].

Against closed flagships, V4 Pro remains cheap — roughly a quarter of the per-token cost of comparable Claude/GPT tiers, even at peak hours. Against its own family, the gap to Flash widened: Flash now runs about a third of Pro's price, so the routing decision (default to Flash, escalate to Pro) is worth real money at volume^[2][6].

The V4 Pro vs Flash comparison and the V4 Pro vs GPT-5.5 comparison pages dig into workload-specific math.

The honest segmentation, based on the evidence above^[2][4][6].

The pattern that keeps showing up in testing is hybrid: use Flash for the 90% of traffic that is easy, escalate to Pro for the 10% that is hard. DeepSeek's own benchmark aggregation backs this — Flash sits around 83-88% of Pro's quality at a third of the price^[4].

If you are coming from the preview, the upgrade decision is basically made for you — the API already points at 0813. The remaining question is workload design, not model choice.

• Buy it: agent startups, code-assistant builders, teams running coding agents with a stable prefix (to exploit cache hits), anyone doing hard multi-step reasoning
• Consider it: teams replacing closed coding agents — the open weights + MIT license remove vendor lock-in at near-flagship quality
• Skip it: high-volume, latency-sensitive chat; simple Q&A at scale; teams that need >500 concurrent requests on one key
• Watch it: teams on tight budgets — the peak-hour rates plus reasoning-token billing can surprise