Three months of waiting, and the release is about to land. It has been nearly a quarter since the DeepSeek V4 preview — a stretch in which Anthropic’s Fable 5 traded blows with GPT-5.6 Sol and Kimi K3 forced its way into the first tier. DeepSeek, meanwhile, said only one thing: the V4 official version is coming. That restraint now reads as strategy: while competitors traded launch-event salvos, DeepSeek spent the quarter tuning a commercial mechanism nobody else has dared to ship.
On July 19, multiple outlets confirmed DeepSeek V4 GA gray testing had begun, with full release as early as July 20. This release brings more than performance — it brings the first peak-valley billing in DeepSeek’s history. The price butcher who never raised prices has, for the first time, installed a meter on its own compute.
Two Versions, Two Jobs: Flash for Volume, Pro for Positioning
The official V4 ships in two tiers: DeepSeek V4 Flash, the lightweight value play, and DeepSeek V4 Pro, the high-performance version benchmarked against Opus-class models. Early tester feedback puts V4 Pro near Claude Opus 4.8 overall, with coding closing on GPT-5.6 Sol and agent abilities — including 3D and SVG generation — markedly improved. But some developers note Pro does not lead Flash by as much as expected; the gap is smaller than the positioning implies.
One detail worth tracking: on identical tasks, V4 needs more iteration rounds than Fable 5, meaning inference efficiency still has headroom. Factor in the price, though, and that “weakness” starts looking different — extra rounds at a fraction of the cost can still be the cheaper answer. The practical takeaway is to measure your own tasks in dollars-per-completed-job rather than rounds-per-task: a model that needs three tries at a seventh of the price beats a model that needs one try at five times the price, unless your bottleneck is latency or human review capacity rather than spend.
Peak-Valley Billing: the New Algorithm of the Price Butcher
This is DeepSeek’s first time-tiered pricing — and it reads as the marker of AI commercialization entering fine-grained operations. The numbers:
-
V4 Pro: $0.87 per million output tokens off-peak, $1.74 at peak; cache-miss input $0.435 off-peak.
-
V4 Flash: $0.28 per million output off-peak, $0.56 at peak; cache-hit input as low as $0.0028 per million tokens.
For scale: Fable 5 charges $50 per million output tokens, and GPT-5.6 Sol more. Even at peak pricing, V4 Pro runs tens of times cheaper. The cache-hit scenario is almost absurdly cheap — irrelevant for individuals, but for teams running agents continuously through business hours, the cost structure demands recalculation. The good news: batch jobs, evaluation runs, and data generation can move off-peak, and the bill compresses right back down.
The Niche: Not First, Just the Best Fit
On SWE-bench Verified, DeepSeek V4-Pro-Max sits 0.2 percentage points behind Claude Opus 4.6 Max. Opus keeps the lead in long-context retrieval, professional software engineering, and parts of knowledge reasoning — but the gap has shrunk to near-imperceptible. Kimi K3 is considered slightly stronger overall, yet DeepSeek’s play is categorically different: performance is not first; value is first. That is the path DeepSeek has validated since V2 — take top-tier capability and price it at a seventh or less.
The industry signal inside peak-valley billing deserves emphasis: it means DeepSeek has recognized that “infinitely cheap” is unsustainable, and instead of raising prices it chose fine-grained incentives to steer user behavior. For the model market, that is a real turning point — the shift from competing on parameters to competing on operations may arrive faster than anyone expected. Watch whether rivals copy the mechanism: pricing structures, once introduced, tend to propagate across a market faster than capability features do, because they are visible in every procurement spreadsheet.
Legacy Models Retire, the New Order Settles
With the V4 official release, the old model names deepseek-chat and deepseek-reasoner go offline on July 24. The last light of the old era switches off. For developers this is an unambiguous instruction: DeepSeek is moving wholesale to the V4 architecture, and legacy API calls need migration now, not later. Migration in this window is mercifully simple — the API surface is compatible, so in most codebases the change is a model-name string plus a re-run of your evaluation suite against V4’s slightly different response character. Do it this week while both generations still respond, and you keep a rollback path; do it after the 24th and any regression debugging happens blind.
Your Move: How to Handle the Release
-
Spot delivery or futures? Gray users can already try V4; full rollout lands around July 20-22. Heavy API users should watch DeepSeek’s official announcements daily this week.
-
Rerun the math. If your agents hit the API at high concurrency during business hours, peak-valley billing means rescheduling: move non-realtime tasks to evening batches and the cost can halve.
-
Test the migration now. The legacy models die in under a week. Still on deepseek-chat or deepseek-reasoner? Switch to the V4 preview immediately and run compatibility tests.
-
Don’t shop by parameter count. V4 is not the strongest model on paper. But if your workload is latency-insensitive — non-realtime conversation, batch processing, offline analysis — it may be the best value available today.
From leaderboard domination to value domination, DeepSeek’s choice carries a signal bigger than itself: as model capability converges, the war’s main front is moving from the rankings to the cost structure. The teams that learn to schedule around peak-valley pricing this month will be the ones whose agent economics survive the year. The rest will keep paying peak rates for jobs that could have run at midnight — a pure operations tax on not reading the email DeepSeek sent everyone.
