By Stuart Kerr, Technology Correspondent, LiveAIWire
DeepSeek V4 pricing crashed the cost of running a frontier-adjacent AI model to a fraction of what Claude or Gemini charge, and it arrived with a hard deadline attached: the two legacy model names that a huge share of existing API calls still point to, deepseek-chat and deepseek-reasoner, were fully retired at 15:59 UTC on July 24, 2026, and any request still using them now fails outright rather than quietly rerouting to a newer model.
DeepSeek’s own API documentation confirms the cutoff in plain terms, and pairs it with something the company has never done before: peak and off-peak pricing that doubles the cost of a request during Beijing business hours. Together, the deadline and the new pricing model mark the moment DeepSeek V4 stopped being a preview developers could experiment with at their own pace and became the production default, on DeepSeek’s own schedule rather than each developer’s.
What Actually Changed When DeepSeek V4 Reached General Availability
DeepSeek’s V4 family first shipped as a preview on April 24, 2026, in two variants. V4-Pro is a mixture-of-experts model built around 1.6 trillion total parameters, of which only 49 billion activate for any given token, a design that keeps the cost of an individual request close to that of a far smaller dense model while retaining far more stored capacity. V4-Flash carries 284 billion total parameters with 13 billion active, built specifically for latency-sensitive, high-volume workloads rather than maximum raw capability.
Both variants support a 1-million-token context window as the default across DeepSeek’s official services, and both remained available with open weights throughout the roughly three-month preview period, a departure from how many rivals stage a model behind a private waitlist before opening it up.
General availability, confirmed in DeepSeek’s own API documentation, folded in agentic capability improvements and stronger mathematics and coding reasoning on top of the preview builds, then activated the new peak and off-peak billing structure at the same time. Peak hours run 9 a.m. to noon and 2 p.m. to 6 p.m. Beijing time, and any call placed inside those windows now carries a price multiplier on top of the base rate, an incentive structure borrowed more directly from electricity retailers than from any prior AI pricing model. Cached inputs get a separate, steeper discount again, which matters most to applications that repeatedly reuse long system prompts or shared retrieval context across many calls.
Why the Legacy Cutoff Matters More Than It Looks
Plenty of API providers handle a model retirement by quietly routing old model names to whatever replaced them, which keeps existing code technically running while silently changing its behaviour and cost without anyone necessarily noticing. DeepSeek chose the opposite approach for deepseek-chat and deepseek-reasoner. Once the July 24 cutoff passed, calls to those names do not fall back to V4 automatically. They fail, returning an error rather than a degraded response, which forces every team still calling the legacy names to make an explicit, visible decision about the migration rather than discovering months later that their production behaviour had shifted underneath them.
For most integrations the actual fix is a single line of code, replacing a call to deepseek-chat or deepseek-reasoner with a call to deepseek-v4-flash or deepseek-v4-pro depending on the workload, since those replacement model IDs have been stable since the original April preview. The teams most exposed by the hard cutoff are the ones running high-volume automated calls triggered by infrastructure rather than a person clicking a button, exactly the kind of scheduled or event-driven workload most likely to have missed a documentation notice buried in a preview-to-GA changelog months earlier.
Why DeepSeek V4 Pricing Is Reshaping the Rest of the Market
DeepSeek-V4-Flash costs 14 cents per million input tokens and 28 cents per million output tokens off-peak, a rate that undercuts most of the closed-source market by a wide margin. Anthropic’s own published pricing lists Claude Opus 4.8 at 5 dollars per million input tokens and 25 dollars per million output tokens, roughly 36 times V4-Flash’s input rate and 89 times its output rate, though the two models are aimed at different tiers of task rather than competing head to head on identical work.
Even Anthropic’s smaller Claude Haiku 4.5 lists at 1 dollar input and 5 dollars output per million tokens, still seven times V4-Flash’s input price, and Google’s comparably positioned Gemini 3.5 Flash runs roughly ten times V4-Flash’s input rate according to a July 2026 pricing comparison.
That gap is not a rounding error for anyone running AI at real production volume. A workload processing a billion input and a billion output tokens a month costs roughly 30,000 dollars on Claude Opus 4.8’s published rate card and roughly 420 dollars on DeepSeek V4-Flash off-peak pricing, a difference of about 71 times on the identical volume. Tasks that were previously too expensive to automate through a frontier-priced model at real scale, bulk document summarisation, first-pass customer message triage, become financially trivial at that price point, which is precisely why enterprise procurement teams are tracking DeepSeek’s general availability rollout as closely as individual developers are.
Why This Is a Quieter Moment Than DeepSeek’s Last Major Release
DeepSeek’s V4 general availability is a markedly quieter event than the one that first put the company on the map. DeepSeek’s R1 model, released in January 2025, triggered a sharp single-day selloff in AI infrastructure stocks as markets abruptly recalculated what frontier-level AI was supposed to cost to train and run, a shock that rippled well beyond the AI industry itself and prompted several governments, including Australia’s, to restrict DeepSeek’s consumer products on official devices over data and security concerns. Eighteen months on, V4’s general availability has followed a similar script, open weights, a steep price cut against the nearest closed-source equivalent, no marketing campaign, without producing anything close to the same market reaction.
That absence of reaction is arguably the more interesting story than the release itself. An open-weight Chinese lab shipping a frontier-adjacent model at a steep discount to Western rivals has become routine enough to barely register outside specialist AI coverage, a sign the market has priced in DeepSeek’s release cadence in a way it had not at the time of R1’s debut.
LiveAIWire’s own coverage of Kimi K3’s open weight release from Moonshot AI found a related pattern the same week, another Chinese lab racing to close the gap with US frontier systems, though Moonshot has chosen to compete on raw scale, a 2.8 trillion parameter model, rather than DeepSeek’s approach of undercutting on price.
What This Means for Anyone Choosing a Model
For a team evaluating DeepSeek V4 against Western alternatives, the honest read on the evidence is that open-weight AI has not closed the capability gap at the very top of the market. Independent comparisons in mid-2026 continue to place Claude Opus 4.8 and comparable frontier systems ahead on the hardest all-round benchmarks, and DeepSeek does not currently sit at the top of those specific leaderboards.
What has changed is that gap’s relevance to the bulk of real production workloads, since a model with a million-token context window and open weights at a small fraction of frontier pricing is genuinely sufficient for the summarisation, classification, and first-draft coding tasks that make up most of what businesses automate, a distinction LiveAIWire’s broader comparison of ChatGPT, Gemini and Claude in 2026 found holding across the industry, that no single assistant wins every task and the smartest approach is matching the tool to the job.
Not every organisation can make that tradeoff on price alone. LiveAIWire’s coverage of the AI Cold War between the United States and China found that chip and infrastructure access between the two countries can shift within months rather than years, and the same geopolitical caution increasingly extends to the models themselves.
Government agencies and regulated businesses in several countries, Australia’s federal government among them, have restricted DeepSeek’s products specifically over data-handling and provenance concerns that a favourable price comparison does not resolve, which means the practical decision facing most teams is not simply whether V4 is good enough and cheap enough, but whether routing data through a Chinese-developed model creates a compliance problem a client or regulator will not accept regardless of the savings on offer.
The Broader Pattern This Fits
DeepSeek’s pricing move is part of a wider trend LiveAIWire has tracked across open-weight AI this year. Our coverage of Switzerland’s fully open Apertus model found a different, government-funded version of the same underlying logic, that relying entirely on AI infrastructure built and controlled by a single foreign company creates a strategic dependency institutions are increasingly willing to pay a real quality cost to avoid, whether that dependency runs through Beijing, Washington, or any single commercial lab. DeepSeek’s steep discount and hard-deadline migration this week is a commercial version of the same pressure, forcing every developer still calling its legacy API names to make an active choice about which model, and which country’s infrastructure, their production system will depend on going forward.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.