DeepSeek Releases Lightweight New Model With Per-Task Costs One-Tenth of US Rivals
Chinese AI company DeepSeek released its next-generation AI model "DeepSeek-V4.1-Flash-1017" on October 17, claiming that with just 16 billion active parameters it surpasses Anthropic's flagship Claude Opus 5 across multiple agentic task benchmarks—at a per-task running cost of less than one-tenth that of comparable US products. The launch comes at a critical juncture as the global large-model industry pivots from "capability competition" to "profit competition," with the low-cost, high-efficiency technical approach once again drawing market attention.
The model is a fine-tuned upgrade of DeepSeek-V4.1-Flash released in September. Model weights have been made publicly available under an MIT license on platforms including Hugging Face, and the "deepseek-flash" model on DeepSeek's official website, mobile app, and API endpoints has been switched to the new version.
A Big Breakthrough for a Small Model
From a technical architecture standpoint, V4.1-Flash employs a Mixture-of-Experts (MoE) design with a total parameter count of 552 billion, plus a 196-billion-parameter "Engram" memory module. However, when actually processing each token, only 8 billion parameters are activated during the input phase and 16 billion during the output phase. This "large storage, small activation" design allows the model to draw on massive knowledge reserves while keeping computational consumption relatively low.
According to benchmark data published by DeepSeek, the model scored 77.0 on DeepSWE v1.1—a test measuring software defect repair capability—beating Claude Opus 5's 74.0 and OpenAI GPT-5.6 Sol's 73.0. On Automation-Bench, which measures business automation proficiency, V4.1-Flash led significantly with 60.5 versus Opus 5's 50.3. DeepSeek claims it surpassed Opus 5 across all four agentic benchmark tests.
However, all of these results come from DeepSeek's own evaluations, with no third-party independent verification yet available.
A Stark Cost Advantage
What drew even more industry attention is the cost gap. According to data cited by German tech outlet The Decoder, on DeepSWE tasks V4.1-Flash costs roughly $0.20 per task at maximum settings and about $0.10 at lower settings—while still maintaining over 72% accuracy at the lower setting. By comparison, competitors such as GPT-6 Astra and Claude Opus 5 cost between $2 and $10 or more per task.
On API pricing, DeepSeek has kept its existing rates unchanged: $0.15 to $0.30 per million input tokens (cache miss) and $0.60 to $1.20 for output tokens. Peak hours are 10:00–13:00 and 15:00–19:00 Japan time on weekdays, with prices halved at all other times.
DeepSeek also opened its Search and Fetch APIs—built on its proprietary search engine—to developers, charging $0.75 per 1,000 successful requests.
The Industry Faces a Profit Reckoning
DeepSeek's strategy of entering the market with extreme cost-performance has struck the industry's most sensitive nerve. Just one day before the new model launch, Anthropic disclosed its latest financials to prospective IPO investors in San Francisco: third-quarter 2026 revenue of $16.8 billion, up 45% quarter-over-quarter, with its first-ever "adjusted profitability" and positive cash flow.
Yet these impressive figures have not fully dispelled market doubts. According to four people familiar with the matter, at the October 16 investor meeting, some existing shareholders and prospective investors are reassessing Anthropic's valuation, with figures around $1.5 trillion under discussion—a notable reduction from the previously anticipated $2 trillion IPO valuation.
| Metric | FY2025 | Q1 2026 | Q2 2026 | Q3 2026 |
|---|---|---|---|---|
| Revenue | approximately $4.6 billion | $4.7 billion | $11.6 billion | $16.8 billion |
| QoQ Growth | — | — | ~147% | ~45% |
| Operating Profit | Loss of $8.06 billion | Not disclosed | Not disclosed | Adjusted profit |
Note: Data from Anthropic's investor disclosures and previously public information.
Decelerating growth is the most immediate source of investor concern. Second-quarter revenue more than doubled sequentially, but third-quarter growth has slowed to roughly 45%. Some analysts had previously projected Anthropic could reach an annualized revenue run rate of $110 billion by year-end—a target that now appears difficult to achieve.
Aswath Damodaran, a professor at NYU Stern School of Business, ran a reverse valuation exercise on Anthropic: assuming the company can eventually achieve a 30% after-tax operating margin and a cost of capital around 10%, a $2 trillion IPO valuation would require annual revenue approaching $1.2 trillion within 10 years of reaching maturity; if it takes 15 years, the figure would need to approach $2 trillion.
The Financial Reality for Chinese Peers
The predicament of Chinese large-model companies reflects the industry's deeper contradictions. Zhipu reported first-half 2026 revenue of CNY 954 million (approximately $142.7 million), up nearly 400% year-over-year, including API revenue of CNY 825 million (approximately $123.4 million), a more than 27-fold increase. Yet net losses for the same period still exceeded CNY 2 billion (approximately $299.1 million), and overall gross margin fell from 50% a year earlier to 26.4%. As of October 14, Zhipu's total market capitalization on the Hong Kong Stock Exchange stood at approximately HK$279 billion (approximately $35.6 billion), down more than 70% from the HK$1 trillion level briefly breached in June.
MiniMax posted first-half revenue of $117 million, up 283% year-over-year, with gross margin improving from 12.1% to 17.9%; however, adjusted net losses widened to $293 million. Its market capitalization on October 14 was only about HK$71.3 billion (approximately $9.1 billion).
The pricing divergence between primary and secondary markets is also becoming increasingly pronounced. According to reports, Moonshot AI's annualized recurring revenue (ARR) surpassed $1 billion in August, and its ongoing pre-IPO financing round values the company at approximately $50 billion—roughly 50 times ARR. DeepSeek's annualized revenue is similarly around $1 billion, with its new funding round targeting a valuation of CNY 500 billion (approximately $74.8 billion), corresponding to about 75 times ARR. Meanwhile, in public markets, Zhipu and MiniMax's valuations have already corrected sharply.
GSR Ventures managing partner Zhu Xiaohu put it bluntly: AI was previously valued on narrative; going forward, it must shift to fundamental-based valuation.
Hardware Players Watch From the Sidelines
Even one of the biggest beneficiaries of the AI boom is sounding warnings. Broadcom CEO Hock Tan, speaking at the DBS Asian Insights Conference in Singapore on October 13, compared many investors pouring into AI to gamblers.
"A lot of investors don't really know what they're investing in," he said. "They set very extreme expectations, and then they find out it's almost like being in a casino."
Broadcom has generated roughly $68 billion in revenue from AI semiconductor business since late 2023 and is a leading designer of custom AI chips for hyperscale data centers. But Tan pointed out that chip suppliers can profit from selling hardware, while AI labs purchasing compute still need to find enough paying customers to cover the costs of developing and running models. OpenAI alone, for example, is expected to spend more than $25 billion on training this year.
"Right now they're still losing money because the training investment is enormous," Tan said. "Frankly, at this stage, it proves nothing other than that frontier generative AI models have massive potential revenue growth."
He also criticized enterprises for blindly investing in AI applications, likening it to "a kid with tokens at an arcade game, spending $10 in 10 minutes and learning nothing."
These warnings echo findings from a recent McKinsey survey: while 80% of respondents said AI improved their personal productivity, only 37% believed AI had a positive contribution to their organization's EBITDA.
Will the "DeepSeek Shock" Repeat?
In January 2025, DeepSeek launched a high-performance model at a training cost far below its US counterparts, triggering sharp volatility in global AI stocks—an episode dubbed the "DeepSeek Shock." This time, V4.1-Flash takes on Anthropic and OpenAI's flagship products with 16 billion active parameters and per-task costs one-tenth of the competition, once again thrusting the "efficiency versus scale" debate into the spotlight.
But a sober assessment is warranted. DeepSeek's published comparisons are limited to four agentic benchmark tests; in broader benchmarks, Claude Opus 5 still holds a clear advantage. Moreover, all data comes from DeepSeek's own evaluations, with no third-party replication yet available.
The broader backdrop is that the US Federal Trade Commission's investigation into frontier AI companies is ongoing, affecting IPO timelines and valuation expectations for leading players such as OpenAI and Anthropic. In an era of retreating capital and profit-first priorities, whether DeepSeek's low-cost approach can truly shake the market position of US giants depends on the reliability of its models in real-world application scenarios and whether third-party evaluations can corroborate its claimed performance.
Over the past three years, the most important question in the large-model industry was "whose model is stronger." Going forward, the question is becoming: who can sell intelligence at a higher price while driving compute costs lower. DeepSeek has clearly gone further down the second path.