After nearly three months of waiting, the DeepSeek V4 official release is finally landing.
Per Synced’s reporting, DeepSeek V4 (GA) gray-test access has opened to selected users, with full launch as early as July 21. The release ships in two versions: DeepSeek V4 Flash and DeepSeek V4 Pro. With GPT-5.6 Sol, Fable 5, and Kimi K3 all having taken the stage, the open-source camp is finally showing its hand.
Performance: Not First Everywhere — But Strong Enough to Fight
Gray-test feedback makes V4’s positioning clear: overall capability near Opus 4.8, coding closing on GPT-5.6 Sol, substantially enhanced agent abilities, and marked improvements in 3D and SVG generation. The weakness is iteration count — the same tasks take V4 more rounds than Fable 5.
Developer Pankaj Kumar’s summary is fair: by niche, V4 probably cannot beat the freshly launched Kimi K3 — but it will price significantly lower. On SWE-bench Verified, DeepSeek-V4-Pro-Max sits just 0.2 percentage points behind Claude Opus 4.6 Max at one-seventh the price, though Opus keeps the lead in long-context retrieval, professional software engineering, and parts of knowledge reasoning.
A charming gray-test shibboleth: if the model’s chain of thought opens with “I’m” or “I’ll” rather than the old “Let me,” you are probably already on V4 GA. That first-person shift hints at a bottom-layer rework of how the model reasons — small tells like this have become how the community fingerprints Gray builds, since release notes rarely capture the texture of a changed model. If your prompts depend on parsing chain-of-thought formatting, verify your parsers against the new voice before shipping.
Pricing: Peak-Valley Billing — the Price Butcher Gets a Meter
The real headline is the pricing. Late last month, DeepSeek emailed every API user announcing the V4 official launch in mid-July plus a pricing restructure — the first-ever peak/valley billing:
-
DeepSeek V4 Pro: $0.87 per million output tokens off-peak, $1.74 at peak; cache-miss input $0.435 off-peak.
-
DeepSeek V4 Flash: $0.28 per million output off-peak, $0.56 at peak; cache-hit input just $0.0028.
For contrast: Fable 5 charges $50 per million output tokens. Meter installed and all, V4 remains crushing on value.
Two signals hide in the strategy. First, DeepSeek is now confronting its own cost structure. The never-raises-prices “price butcher” installed a meter for the first time. Little effect on individuals — but teams running agents nonstop through business hours just inherited a real number to recompute. Second, the absurdly cheap cache-hit price ($0.0028 per million tokens) is behavioral steering: move batch jobs, evaluation runs, and data generation outside peak hours and the cost compresses further. DeepSeek is teaching its users to schedule.
The Frame: DeepSeek’s Niche Strategy
V4 will probably not be first on every benchmark. Fable 5 and GPT-5.6 Sol hold the high end, Kimi K3 attacks the flank, and winning on raw performance is a hard hand. But DeepSeek never plays that card. Its strategy is a repeatedly validated old path: take Opus-class capability and price it at one-seventh. That is not merely a price war — it is niche competition: with performance good enough, value opens the broader application-layer market.
The lineage tells the story. DeepSeek V2 proved the value MoE model was viable; V3 and R1 broke out on reasoning and detonated the “DeepSeek moment”; V4 pivots from technical breakthrough to commercial operation, peak-valley billing included. The essence: once capability reaches Opus level, the contest shifts from “who is stronger” to “who is cheaper, more reliable, and easier to deploy at scale.” DeepSeek is completing its transformation from lab star to commercial infrastructure.
What It Means for the Application Layer
Three consequences for AI application developers. 1. The cost threshold for commercial agents drops again. At $0.28 per million output tokens and below, token spend stops being the bottleneck for agent chains previously shelved over inference costs — time to re-evaluate them. 2. Peak-valley billing creates a new engineering practice: usage-pattern optimization. Not every task belongs in peak hours; batch jobs, automated evals, and data generation should be scheduled into the valleys. 3. Open-model competitiveness lives in usability, not benchmark scores. Gray feedback says V4 is not the best — but the “good enough plus cheap” combination covers most application scenarios, which means application-layer competition shifts toward product experience and business fit rather than waiting for the next capability jump.
There is also a competitive second-order effect worth naming. Peak-valley billing only works if DeepSeek’s valleys stay empty enough to be cheap — which depends on the market having alternatives to fill its peaks. Rivals priced uniformly across hours now face an awkward choice: adopt the same mechanism and legitimize it, or hold flat pricing and quietly subsidize DeepSeek’s off-peak economics. Pricing mechanisms, once introduced, tend to propagate — and whoever introduces them first also gets to shape what “peak” means.
Action Items
-
Get onto V4 during gray testing and evaluate it on your actual workloads, not on leaderboards.
-
Recompute your agent token costs, especially for high-concurrency, long-chain agent scenarios.
-
Build a cache strategy to capture the $0.0028 cache-hit price, and schedule batch work off-peak.
-
Watch July 24: the deepseek-chat and deepseek-reasoner model names go offline — migrate early.
A side note: DeepSeek’s valuation recently leaked through a luggage company’s disclosure — oblique confirmation of its commercial value. But the pricing strategy is the more meaningful signal. It says DeepSeek is walking from “technical star” to “commercial infrastructure,” and that road — not a leaderboard — is what a real DeepSeek moment looks like.
References: Synced (Zhiyuan Community) reporting · Pankaj Kumar’s gray-test feedback · DeepSeek’s API user email announcement
