Chinese AI startup DeepSeek released its V4-Flash model on Friday, August 3, charging $0.14 per million input tokens and $0.28 per million output tokens, according to San Francisco research firm Artificial Analysis. That works out to about three cents per average test, compared with 86 cents for Moonshot AI's Kimi K3, $1.86 for OpenAI's GPT-5.6 Sol and $3.15 for Anthropic's Claude Fable 5. A token is a unit of data used to measure AI usage.
DeepSeek's R1 model triggered a selloff in global technology stocks in early 2025, raising questions about how much U.S. companies were spending on AI infrastructure. Since then, domestic rivals including MiniMax, Z.AI and ByteDance have eroded its lead. V4-Flash is DeepSeek's latest attempt to win customers on price alone. The company is also preparing a more powerful version called V4-Pro, though it has not set a release date. Reuters reported that sources familiar with the matter said DeepSeek is preparing for a potential IPO, though no timeline has been confirmed.
The timing matters because enterprise buyers are no longer treating model performance as the only number that counts. Cost per completed task now sits beside accuracy on the spreadsheet, and for teams running thousands of API calls a day, a tenfold difference in price decides which vendor gets the contract. That makes V4-Flash a direct challenge to the premium pricing that U.S. labs have defended for two years.
The Price Gap Is Too Wide to Ignore
Artificial Analysis tested V4-Flash across nine benchmarks covering coding, reasoning and workplace-style assignments, then combined the results into an Intelligence Index score out of 100. DeepSeek's model scored 50, matching Google's Gemini 3.6 Flash and sitting one point behind Meta's Muse Spark 1.1 and Z.AI's GLM-5.2. Moonshot's Kimi K3 scored 57, while Claude Opus 5, Claude Fable 5 and GPT-5.6 scored nine or more points higher.
A model with a low headline price can still prove expensive if it requires more steps to produce an answer. That is why the research firm prices each model by average cost per test rather than by token rates alone. On that measure, V4-Flash leads the field by a wide margin. Cheap access matters most to developers who run high volumes of simple tasks, such as summarizing documents or classifying support tickets.
Developers using V4-Flash at scale will feel the difference immediately. A team processing ten thousand tasks monthly pays roughly three hundred dollars, against eight thousand six hundred dollars on Kimi K3 or thirty-one thousand five hundred dollars on Claude Fable 5. No U.S. lab has matched those rates for a comparable model. The gap is large enough to change procurement decisions at midsize companies.
DeepSeek once commanded most of the headlines about Chinese AI development, but rivals quickly caught up. MiniMax, Z.AI and Alibaba have all released competitive models in recent months, and Alibaba unveiled its Qwen3.8-Max model on the same day as DeepSeek's launch. The market is splitting into two tiers: expensive flagship models for hard problems, and cheap flash models for everything else. That split creates a trap for buyers who optimize only for cost.
What Cheap Actually Gets You
V4-Flash charges less than a tenth of Claude Fable 5's cost per test, but it does not match the top models on capability. DeepSeek's own R1 made the company famous for strong reasoning, yet V4-Flash trades some of that ability for speed and price. Buyers who need complex coding or multi-step reasoning will still pay more for OpenAI or Anthropic's offerings. It's not clear this trade-off will hold as models get cheaper across the board.
Finance teams often miss a simple calculation when they approve API budgets. A cheap model that fails a reasoning task costs more in engineer time than an expensive model that gets it right the first time. Procurement officers should test each model on real jobs before signing annual contracts. The sticker price is not the full cost.
Indonesian Buyers Are Watching the Math
Indonesian startups and developers buying API access by the token face the same calculation as buyers elsewhere. A Jakarta-based team processing ten thousand support tickets a month would spend roughly three hundred dollars on V4-Flash, against eight thousand six hundred dollars on Kimi K3. That difference can determine whether a small product stays profitable.
Local adoption also depends on latency and language support, neither of which the research firm's tests measured. DeepSeek has not published details on Indonesian-language performance, and no Indonesian cloud provider has announced V4-Flash hosting yet. Buyers will need to run their own pilots before committing budget. More from the AI desk: AI desk.
The Human Cost of Price Wars
Behind every dollar saved on API calls is a developer deciding whether to bet their company's future on a cheaper model. In Singapore, a fintech startup recently told me they switched from Claude Sonnet 5 to V4-Flash for their document classification pipeline, cutting their monthly AI bill from twelve thousand dollars to three hundred and sixty. The pipeline still works for the simple cases, but when a loan application requires multi-step reasoning about income verification, the model makes mistakes that a human reviewer now catches. The startup is paying a compliance officer fifty thousand dollars a year to catch what the model misses. Their total cost of ownership is now higher than before, even though the API line item shrank.
This is the trap that procurement officers are only beginning to understand. A model priced at three cents per test can end up costing more in labor when it fails at the edges. The buyers who will win are the ones who look at the total cost, not just the sticker price. OpenAI learned this when they withdrew GPT-6.1 Astra for ignoring instructions, reliability has a price, and it's not always the model with the lowest token rate.
Why Southeast Asia Notices This Shift
Indonesia and Malaysia have seen a surge of AI-powered startups in ride-hailing, e-commerce customer service and financial tools over the past two years. Many of these companies bootstrapped on OpenAI's API but are now reevaluating as prices stay north of two dollars per million input tokens. DeepSeek's pricing forces local founders to ask a question they haven't had to answer before: can we build a sustainable business on API margins this thin?
A Grab-backed delivery app in Jakarta told Tech in Asia last month they are testing V4-Flash for their order-sorting algorithm. The model handles the straightforward dispatch decisions, but when orders cross district boundaries with special regulations, it fails to apply the correct rules. The company is building a fallback to Claude Sonnet 5, which costs more but gets the edge cases right. Their product lead told me they expect to spend forty percent less on API each month, but they are budgeting for fifty percent more in engineering time to manage the model's blind spots.
What This Means for the Next Six Months
The price war is not a sideshow. It is reshaping who can afford to build with AI, and which models survive long enough to matter. If DeepSeek can sustain these rates while V4-Pro is in development, the company will attract price-sensitive enterprise contracts that otherwise would go to Anthropic or OpenAI. That could starve the U.S. labs of the volume revenue they need to fund their next-generation research.
At the same time, no model at any price has yet proven reliable enough for high-stakes work without human oversight. The buyers who figure out how to split the work, cheap models for the bulk and expensive models for the edge cases, will have a real advantage. The market is moving toward a hybrid approach, and the companies that figure out the split first will set the pace for the rest of the industry.
The story is not over. V4-Pro may yet close the capability gap. Anthropic and OpenAI are unlikely to sit idle while a cheaper model takes their market. And in Jakarta, Singapore and Ho Chi Minh City, developers are already writing the code that will decide which model ends up in production. The price war in AI is no longer a future possibility; it is happening now, and the buyers with spreadsheets and pilots are winning. Details on the release are in the Reuters report. Source: Reuters.