1. V4 Pro Goes Official — The V4 Family Is Complete
In the early hours of August 13, 2026, DeepSeek officially promoted V4 Pro from preview to production, updating the model version to DeepSeek-V4-Pro-0813 and making it available through the API. With this release, both versions of the V4 family — Flash (launched late July) and Pro — have now reached official status.
The release was pre-announced: DeepSeek's API changelog back in April flagged that a formal V4 Pro launch would follow. The official pricing page now carries the production model version, and developers can switch by changing the model parameter to deepseek-v4-pro — the base URL and API interface remain unchanged.
The timing is strategic. V4 Flash's official version (July 31) had just topped the global OpenRouter token usage rankings, demonstrating massive adoption. The Pro launch completes the two-tier product structure: Flash for high-frequency, high-concurrency, cost-sensitive workloads (concurrency limit 2,500); Pro for high-specification tasks (concurrency limit 500).
2. The Agent Leap: From 7.3 to 62.7
The headline improvement in V4 Pro 0813 is agent capability. Compared to the V4-Pro-Preview, the production model shows dramatic gains across agent-focused benchmarks:
| Benchmark | Preview | V4-Pro-0813 | Comparison |
|---|---|---|---|
| DeepSWE (AI coding agent) | 7.3 | 62.7 | Surpasses Claude Opus 4.8 |
| Cybergym (AI security agent) | — | Leading | Surpasses Claude Fable 5 |
| AutomationBench (agent workflow) | — | Leading | Surpasses Claude Fable 5 |
The DeepSWE jump — from 7.3 to 62.7 — is particularly significant. DeepSWE measures an AI agent's ability to independently resolve software engineering tasks: reproduce the issue, locate the bug, implement the fix, and run tests. A near-10x improvement indicates the model is no longer just generating code snippets but operating as a genuinely autonomous coding agent.
The benchmark results position V4 Pro 0813 alongside — and in agent-specific tasks, above — the strongest Western models. This matters because the AI agent is increasingly the deployment layer for AI adoption: enterprises don't consume models directly, they consume agents built on models.
3. The Price Structure: Pro = 3x Flash, Plus an Across-the-Board Hike
V4 Pro's official pricing confirms its premium positioning — and carries an important warning:
| Pricing (per million tokens) | V4 Pro | V4 Flash | Ratio |
|---|---|---|---|
| Input (cache hit) | 0.025 yuan | 0.02 yuan | 1.25x |
| Input (cache miss) | 3 yuan | 1 yuan | 3x |
| Output | 6 yuan | 2 yuan | 3x |
The critical note: DeepSeek's pricing page now carries an official notice — "plans to raise overall DeepSeek API service pricing in the near term, with an expected substantial increase. Please arrange your usage accordingly. The specific plan will be announced via official notice."
This is a significant signal for the Chinese AI industry. DeepSeek has been the price-disruption benchmark — its low pricing drove down costs across the market. A substantial across-the-board price increase would mark a strategic shift: from growth-at-any-cost pricing to monetization phase. For heavy API users, the implication is immediate: pre-purchase credits while current pricing lasts, and reassess long-term cost structures.
4. Implications for Brands and Developers
For TMG's audience — cross-border marketers and the brands they serve — the V4 Pro launch carries four implications:
Implication 1: The economics of AI-powered marketing are changing.
DeepSeek's price increase will ripple through the Chinese AI ecosystem. Brands using DeepSeek for content generation, translation, or data processing should re-budget for higher API costs — and watch whether other providers follow suit. The era of effectively free AI compute is ending.
Implication 2: Agent-grade models change what AI can do for brands.
V4 Pro's agent benchmarks mean the deployment layer is shifting from "models that answer" to "agents that act." For brands, this accelerates the scenarios we've covered — WorkBuddy-style document agents, WeCom-style communication agents, AI video production pipelines. The question is no longer whether AI can do the work, but how brands structure the workflow.
Implication 3: Model selection is now a cost-structure decision.
The Flash-Pro tiering — and the 3x price gap — means model selection is now a meaningful cost-structure decision. The July analysis of "DeepSeek V4 Flash's peak-valley pricing" already established the pattern; the Pro launch completes it. Brands should map their use cases to the right tier: high-volume simple tasks to Flash, complex agent workflows to Pro.
Implication 4: Chinese model providers are entering monetization phase.
DeepSeek's price hike signals that Chinese AI infrastructure is shifting from subsidized growth to sustainable economics. This is healthy for the industry long-term — but it means brands should treat current API pricing as transitional, and build cost contingencies into AI-dependent operations.
Key Takeaways
- DeepSeek V4 Pro went official Aug 13 — model version DeepSeek-V4-Pro-0813, completing the V4 family (Flash + Pro)
- Agent capability leap: DeepSWE 7.3 → 62.7 (surpasses Claude Opus 4.8); Cybergym & AutomationBench surpass Claude Fable 5
- Specs: 1M context, 384K max output, thinking/non-thinking, Tool Calls, Responses API, Anthropic API compatible
- Pricing: Pro = 3x Flash (3 yuan input / 6 yuan output per M tokens); concurrency 500 vs Flash's 2,500
- Official price hike notice: "substantial increase" expected — heavy API users should pre-purchase credits
- For brands: re-budget AI costs, expect agent-grade capabilities, map use cases to Flash/Pro tiers, treat pricing as transitional