DeepSeek
DeepSeek is a Chinese AI chatbot and large language model platform offering free conversational AI, code generation, and reasoning capabilities. Built by a subsidiary of hedge fund High-Flyer Capital Management, it open-sources its models under the MIT license while routing all user data through servers in China.
Score generated by AI agents based on publicly cited evidence and reviewed by the project maintainer. Not independently validated. Last assessed 2026-10-25. Score revised 2026-10-25: 28 → 35.
Score History
Timeline events are AI-curated from public reporting. Score trajectory is derived from documented events.
DeepSeek was spun off from hedge fund High-Flyer on July 17, 2023 as an AGI research lab financed by its parent, with founder Liang Wenfeng holding about 84%. It released open code and language models from November 2023 under permissive licenses and had no mass-market consumer product. China's generative AI rules, in force from August 2023, set the censorship baseline, and control sat entirely with the founder.
DeepSeek V2's release on May 6, 2024, at RMB 1 per million input tokens, set off a Chinese LLM price war, and context caching in August cut cached-input costs by up to 90%. Developers got cheap, commercially usable models, while the chatbot carried state-mandated political filtering that drew little attention outside China. The December 2024 V3 release put a frontier-class open-weights model in the hands of a founder-controlled lab with growing state ties.
The R1 launch on January 20, 2025 took DeepSeek's app to No. 1 on the U.S. App Store and erased about $589 billion of Nvidia's market value in a day. Scrutiny followed for more than a year: the Wiz database exposure, Italy's processing limitation, government-device bans from NASA to the Czech Republic, Korea's findings of unconsented transfers, Germany's app-store notice, Texas's investigation, the House committee's 'profound threat' report and NIST's CAISI evaluation. Censorship was documented and tightened with R1-0528, and distillation accusations from OpenAI and Anthropic piled up into early 2026, while the app stayed free and API prices kept falling.
DeepSeek's first outside funding round, reported June 16, 2026 at more than 50 billion yuan, ended its sole reliance on High-Flyer; investors sit in a founder-managed partnership while a state AI fund holds voting rights. Within months it added peak-hour API surcharges that raised August rates more than fourfold at peak, hired CITIC Securities for a STAR Market IPO and lined up a second round of about 100 billion yuan. Anthropic alleged it silently relayed customers' requests to Claude, China's CAC opened a probe, U.S. agencies named it a top distillation offender and DHS ordered its products out of federal civilian systems. The consumer app remains free and the weights open.
Alternatives
Anthropic's AI assistant, strong at coding, analysis, and long-form content with a 200K token context window. Free tier available, Pro at $20/month. Designed with safety-first principles and no government-mandated censorship. Easy switch with no data migration needed.
The most widely used AI chatbot, with strong reasoning capabilities. Free tier available, Plus at $20/month. Run by a U.S. company with no state-mandated political censorship. Easy switch: just sign up and start using it.
Google's AI chatbot with a 1M token context window and tight integration with Google Workspace. Free tier available, Google AI Pro at $19.99/month. Real-time web grounding reduces hallucinations. Easy switch, especially if you already use Google services.
Dimensional Breakdown
Summaries below were written by AI agents based on the cited evidence. They are editorial interpretations, not independent research findings.
Dimension History
Timeline (59 events)
DeepSeek Spun Off from High-Flyer as Independent Company
High-Flyer Capital Management's AGI research lab, announced in April 2023, was spun off as an independent company called DeepSeek on July 17, 2023, with High-Flyer as its principal investor and backer. Liang Wenfeng, who co-founded High-Flyer and is DeepSeek's CEO, held about 84% of DeepSeek through two shell corporations as of May 2024. Venture capital investors were initially reluctant to fund the lab, so it was financed by High-Flyer rather than outside venture capital. A January 2025 Exovera corporate profile identifies Liang as the controlling shareholder and managing director of Zhejiang High-Flyer Asset Management Company.
China's Generative AI Regulations Mandate Censorship Compliance
China's Interim Measures for the Management of Generative AI Services took effect on August 15, 2023. They require generative AI services offered to the public to uphold the Core Socialist Values and bar content inciting subversion of national sovereignty or the overturn of the socialist system, separatism, or undermining national unity and social stability. The Measures do not cover firms that only research and develop generative AI without offering services to the mainland public, so they applied to DeepSeek (founded weeks earlier) once it offered public services, and they shaped the content controls its chatbot later applied.
DeepSeek Releases First Open-Source Model
DeepSeek released DeepSeek Coder, its first open model, on November 2, 2023, followed by the DeepSeek-LLM 7B/67B series on November 29, 2023. Their code was released under the MIT license and their weights under DeepSeek's own model license, which permits commercial use, establishing DeepSeek's open-weights approach. The models were competitive but not yet frontier-level.
High-Flyer Designated National High-Tech Enterprise
Zhejiang authorities designated Ningbo High-Flyer as a National High-tech Enterprise in December 2023, granting preferential tax policies and state subsidies for R&D. This deepened the connection between DeepSeek's parent company and the Chinese government's industrial policy apparatus.
DeepSeek V2 Triggers Chinese AI Price War
DeepSeek released V2, a 236-billion-parameter model, on May 6, 2024, with API pricing of RMB 1 (about $0.14) per million input tokens and RMB 2 per million output tokens, far below rivals such as Baidu's Wenxin 4.0 at RMB 120. KrASIA described the release as the catalyst for a Chinese LLM price war: Zhipu AI cut GLM-3-Turbo to RMB 1 per million tokens on May 11, ByteDance priced its Doubao model at RMB 0.8 per million tokens on May 15, and on May 21 Alibaba Cloud cut Qwen-Long's price 97% from RMB 20 to RMB 0.5 per million tokens, after which Baidu made two Ernie models free.
DeepSeek Adds Context Caching and Cuts Cached-Input API Costs
DeepSeek introduced disk context caching for its API, charging $0.014 per million tokens for cache hits, which it said cut API costs by up to 90%.
DeepSeek V3 Released as Open-Source Frontier Model
DeepSeek released V3, a 671-billion-parameter mixture-of-experts model trained on 14.8 trillion tokens. Its technical report put the final training run at $5.576 million, 2.788 million H800 GPU hours at an assumed $2 per GPU-hour rental price, excluding prior research and ablation experiments. SemiAnalysis estimated DeepSeek's total server capital expenditure at about $1.6 billion (some coverage reported $1.3 billion). The report described the model as comparable to leading closed-source models, and it was released with open weights. The release put a frontier-level model under the control of a company in which founder Liang Wenfeng held about 84%.
DeepSeek R1 Launch and Premier Li Qiang Meeting
DeepSeek released R1, a 671-billion-parameter reasoning model under the MIT License that it said performs comparably to OpenAI's o1 (see the R1 Launch milestone). On the same day, founder Liang Wenfeng was one of nine individuals asked to speak at Premier Li Qiang's closed-door symposium on the draft Government Work Report, a sign of Beijing's recognition of DeepSeek's place in China's AI ambitions.
DeepSeek App Hits #1, Nvidia Loses $589 Billion
DeepSeek's chatbot app displaced ChatGPT as the No. 1 free app on the U.S. App Store, reaching the top spot on January 26. On January 27, Nvidia stock fell about 17%, erasing about $589 billion in market capitalization, the largest single-day loss for a U.S. company, and the tech-heavy Nasdaq fell 3.1%. Investors wiped roughly $1 trillion off the market value of major U.S. tech stocks in the sell-off.
Large-Scale DDoS Attack Forces Registration Suspension
DeepSeek temporarily limited new user registrations, citing large-scale malicious attacks on its services. Existing users could still log in. The attack was believed to be a DDoS against its API and web chat platform; the company shared no details.
Promptfoo Study Reveals 85% Political Censorship Rate
Promptfoo published a dataset of 1,360 prompts on topics likely to be censored by the Chinese Communist Party and tested DeepSeek-R1 against it. About 85% of the prompts, roughly 1,156, were refused. Topics include Taiwanese independence, historical narratives around the Cultural Revolution, and questions about Xi Jinping. The accompanying dataset also includes Tiananmen Square protests and crackdown prompts, and prompts concerning Uyghur treatment in Xinjiang. The dataset was published on HuggingFace and Google Sheets.
Wiz Discovers Exposed Database with 1M+ Chat Logs
Wiz Research discovered a publicly accessible ClickHouse database at oauth2callback.deepseek.com containing over one million log lines, including user chat histories in plaintext, API keys, secret keys, and backend system details. The database required no authentication, allowing anyone to execute arbitrary SQL queries via a web browser. Wiz disclosed the issue to DeepSeek, which promptly secured the exposure. Wiz CTO Ami Luttwak told Reuters that DeepSeek took the exposed data down in less than an hour after being alerted.
Global Censorship Exposure: AI Deletes Own Answers
International media documented DeepSeek's censorship system in real time, showing the chatbot beginning answers to politically sensitive questions before erasing them and replacing them with a refusal. CNN and NBC News published hands-on tests, and TIME ran an Associated Press comparison of DeepSeek's and ChatGPT's answers. NBC quoted Citizen Lab director Ron Deibert saying it is unusual for China-based applications to censor international users. Freedom House's January 2025 China Media Bulletin also reported that the bot cut off responses mid-sentence, erased generated text and suggested another topic.
OpenAI Says DeepSeek May Have Distilled Its Models
OpenAI told Axios it was aware of and reviewing indications that DeepSeek may have inappropriately distilled its models, noting that its terms forbid using model outputs to build competing models. Trump administration AI czar David Sacks said there was 'substantial evidence' that DeepSeek had distilled knowledge out of OpenAI models and that it was 'possible' intellectual property theft had occurred.
Italy Blocks DeepSeek Over GDPR Violations
Italy's Garante per la protezione dei dati personali ordered an urgent, immediate limitation on the processing of Italian users' data by DeepSeek's two Chinese operating companies. The authority said their reply to its information request was insufficient and rejected their claim that they did not operate in Italy and were not subject to EU rules. It opened an investigation, and DeepSeek was pulled from Italian app stores, though it could still be reached with a VPN. In its January 30 order, the Garante also found that DeepSeek had not clarified its processing activities, that its privacy policy failed GDPR information duties and did not specify legal bases for each processing activity, and that data were stored in China in breach of GDPR safeguards, particularly Article 32.
NASA and U.S. Navy Ban DeepSeek from Systems
NASA blocked DeepSeek from its systems and employee devices on January 31, 2025. The U.S. Navy warned its members against using DeepSeek citing 'potential security and ethical concerns associated with the model's origin and usage.' These were among the first U.S. federal agency bans, preceding broader legislative action.
Cisco Finds 100% Jailbreak Success Rate Against DeepSeek R1
Cisco's Robust Intelligence team, in collaboration with University of Pennsylvania researchers, tested DeepSeek R1 against 50 prompts from the HarmBench dataset. The model failed to block a single harmful prompt, yielding a 100% attack success rate. By comparison, OpenAI's o1 preview had a 26% attack success rate and GPT-4o had 86%. DeepSeek provided cybercrime instructions, misinformation, and harmful content without resistance.
Australia Bans DeepSeek from Government Devices
Australia banned DeepSeek from federal government devices under a Home Affairs direction signed February 4, 2025. Taiwan had already barred government agencies from using the service. In South Korea, the Ministry of Trade, Industry and Energy temporarily banned employees from using DeepSeek on their devices, and the privacy regulator later had the app removed from local app stores pending a review of its data practices.
U.S. Bipartisan Bill to Ban DeepSeek from Federal Devices
Representatives Josh Gottheimer (D-NJ) and Darin LaHood (R-IL) introduced the 'No DeepSeek on Government Devices Act,' which would ban federal employees from using DeepSeek on government-owned electronics and would also cover AI applications developed by its parent, the hedge fund High-Flyer. In May 2025, Senators Bill Cassidy (R-LA) and Jacky Rosen (D-NV) introduced the Protection Against Foreign Adversarial Artificial Intelligence Act, which bars federal contractors from using DeepSeek or High-Flyer applications in performing federal contracts. The formal H.R. 1121 text, dated February 7, lists Gottheimer and LaHood with 16 additional House colleagues. It would require OMB to develop removal standards and guidelines within 60 days after enactment, covering DeepSeek and successor applications or services from High-Flyer or its entities, with national-security and research exceptions.
NowSecure Uncovers iOS App Security and Privacy Flaws
NowSecure published an analysis of DeepSeek's iOS app revealing multiple security and privacy flaws, including unencrypted data transmission, hardcoded encryption keys, and extensive collection of user and device data that can be used for fingerprinting. Separately, SecurityScorecard's STRIKE team reported on February 10, 2025 that DeepSeek's Android app uses anti-debugging mechanisms designed to obstruct security analysis. In a separate February 7 static review of Android version 1.8.0, researcher Michael Bacarella reported code for gathering IMEI identifiers and SIM-card details; he cautioned that static analysis did not prove those functions were executed in practice.
DeepSeek Quietly Removes Keystroke Collection from Privacy Policy
DeepSeek's February 14, 2025 privacy policy update removed the disclosure that it collected 'keystroke patterns or rhythms.' The original policy had listed this behavioral data, which commentators warned could help identify users. The removal raised questions about whether the practice stopped or was merely no longer disclosed. Public criticism preceded the update: on January 29, Tom's Guide warned that VPNs could not protect users against the data collection described in the policy, including its keystroke disclosure. South Korea's PIPC subsequently confirmed that DeepSeek had listed keystroke patterns in its policy but had not collected them.
Texas Attorney General Opens DeepSeek Investigation
Texas Attorney General Ken Paxton announced an investigation into DeepSeek's privacy practices and its claims about its AI model, notified the company that Texas considered its platform to violate the Texas Data Privacy and Security Act, and sent civil investigative demands to Google and Apple seeking their analysis of the app.
DeepSeek Suspends New Downloads in South Korea During Privacy Review
DeepSeek temporarily suspended its service in South Korea from February 15, 2025 to comply with the Personal Information Protection Act, after the PIPC recommended a suspension until necessary fixes were made. The app was removed from the Korean App Store and Google Play; existing users could still access the service.
20+ Central Government SOEs Integrate DeepSeek
Following the central government's launch of an 'AI+' program, at least 20 companies owned by China's central government integrated DeepSeek into their operations. State-owned enterprises including Sinopec, PetroChina, CNOOC, and China Southern Power Grid linked their AI systems to DeepSeek-R1. SASAC met with central SOEs on February 19 to plan future AI work, deepening the structural ties between DeepSeek and the Chinese state apparatus. Separately, Gasgoo reported on February 8 that Dongfeng Motor's MHERO 917 smart cockpit had integrated DeepSeek-R1, with an OTA rollout planned by April 2025.
DeepSeek Introduces Off-Peak API Discounts of Up to 75%
DeepSeek introduced off-peak API discounts from 16:30 to 00:30 GMT daily, during which R1 usage cost 75% less and V3 usage 50% less, according to Reuters.
V3-0324 Moves DeepSeek Model Weights to the MIT License
With the V3-0324 update, DeepSeek said its open-source repositories, including model weights, would use the MIT License, consistent with R1, and that users could use model outputs to train other models, including through distillation.
House Select Committee Report Calls DeepSeek 'Profound Threat'
The bipartisan House Select Committee on the CCP released a report titled 'DeepSeek Unmasked,' calling DeepSeek 'a profound threat to our nation's security.' The report alleged DeepSeek funnels American user data to the CCP, manipulates more than 85% of responses to suppress content about democracy, Taiwan, and human rights, and operates on tens of thousands of Nvidia chips, some of which are subject to U.S. export controls. The committee sent a formal letter to Nvidia demanding answers.
Korea's PIPC Accepts DeepSeek Fixes and Finds Keystroke Patterns Were Not Collected
The PIPC concluded its review and recommended that DeepSeek establish a legal basis for cross-border transfers, delete user input sent to Beijing Volcano Engine Technology, and publish its privacy policy in Korean. DeepSeek had added an opt-out from use of user input for AI training on March 17 and blocked transfers to Volcano on April 10. The PIPC also reported that DeepSeek had listed keystroke patterns in its privacy policy but did not collect them.
South Korea Finds DeepSeek Transferred Data Without Consent
South Korea's Personal Information Protection Commission found that DeepSeek transferred user data to three companies in China and one in the U.S. without consent, and estimated that data from about 1.5 million users may have been improperly transferred during its one-month service period. The investigation also found DeepSeek lacked an opt-out function for AI training, had no age verification despite claiming not to collect children's data, and provided its privacy policy only in Chinese and English.
R1-0528 Update Found to Be Most Censored Model Yet
Testing by the SpeechMap platform found that DeepSeek's R1-0528 update was 'substantially' less permissive of contentious free speech topics than previous releases, earning the designation of 'the most censored DeepSeek model yet for criticism of the Chinese government.' The model censored answers about Xinjiang internment camps and often defaulted to the Chinese government's official stance on sensitive topics.
U.S. Official Says DeepSeek Aids China's Military and Evaded Export Controls
A senior U.S. State Department official told Reuters that DeepSeek has willingly supported China's military and intelligence operations, shares user information and statistics with Beijing's surveillance apparatus, and sought to use Southeast Asian shell companies to obtain advanced chips barred from China. The official said DeepSeek appears more than 150 times in procurement records of the People's Liberation Army and related entities; Reuters could not verify the procurement data and DeepSeek did not respond.
Germany Declares DeepSeek Data Transfers Unlawful
Berlin Data Protection Commissioner Meike Kamp declared DeepSeek's transfer of user data to China unlawful and reported the app to Google and Apple as 'illegal content,' requesting its removal from German app stores. German data protection officials had asked DeepSeek on May 6 to voluntarily remove its apps or fulfill legal requirements for lawful third-country data transfers, but the company failed to comply.
Czech Republic Bans DeepSeek in Government Work
Czech Prime Minister Petr Fiala announced a ban on DeepSeek products across the country's public administration after the national cybersecurity watchdog flagged threats of unauthorized access to users' data. The watchdog warned that DeepSeek is legally obligated to cooperate with Chinese state authorities, creating unacceptable risk for government data.
DeepSeek Releases V3.1 and Announces End of Off-Peak Discounts
DeepSeek released V3.1 with hybrid thinking and non-thinking modes, stronger agent and tool-use skills, and support for the Anthropic API format. It announced that new API pricing would start and off-peak discounts would end at 16:00 UTC on September 5, 2025.
DeepSeek Cuts API Prices by More Than 50% with V3.2-Exp
DeepSeek released V3.2-Exp, built with its DeepSeek Sparse Attention mechanism, and cut its API prices by more than 50%, effective immediately.
NIST's CAISI Finds DeepSeek Models Lag and Echo CCP Narratives
The Center for AI Standards and Innovation at NIST evaluated DeepSeek's R1, R1-0528 and V3.1 against four U.S. models. It found DeepSeek models echoed four times as many inaccurate and misleading CCP narratives as the U.S. models, that R1-0528 answered 94% of overtly malicious requests under a common jailbreak (versus 8% for U.S. models), and that agents built on it were 12 times more likely to be hijacked.
CrowdStrike Finds Political Bias Degrades Code Security
CrowdStrike research revealed that DeepSeek R1 generates code with significantly more security vulnerabilities when prompts mention politically sensitive topics. Coding prompts referencing Tibet increased severe vulnerabilities by nearly 50% (to 27.2%), and mentions of Falun Gong triggered an 'intrinsic kill switch' where the model planned full responses but refused to output code. The behavior appeared embedded in model weights rather than enforced by external guardrails.
DeepSeek Launches V3.2 and V3.2-Speciale
DeepSeek launched V3.2 as the official successor to V3.2-Exp, live on its app, web service and API, with reasoning integrated into tool use. The higher-compute V3.2-Speciale variant was offered through a temporary API endpoint.
Belgium's Federal Government Ban on DeepSeek Takes Effect
Belgian federal public administration employees had until December 1, 2025 to uninstall DeepSeek's apps from work devices, a preventive restriction that followed concerns raised by the Centre for Cybersecurity Belgium.
Reports Allege DeepSeek Trained on Banned Nvidia Blackwell Chips
The Information reported that DeepSeek was using Nvidia Blackwell chips smuggled into China, whose export there is barred by U.S. law, to develop its upcoming model. Nvidia responded that it had not seen any substantiation for the report. Bloomberg's account of The Information's report said the chips were first installed in data centers in countries allowed to buy them, then dismantled and shipped to China. The senior U.S. official's separate Inner Mongolia allegation was reported in February 2026.
Italy's Antitrust Authority Closes DeepSeek Hallucination Probe with Commitments
Italy's AGCM ended its investigation, opened in June 2025, into whether DeepSeek failed to warn users that it may produce false information. It accepted binding commitments from DeepSeek's two operating companies to make disclosures about hallucination risk easier, more transparent and more immediate, without finding an infringement.
OpenAI Warns Congress of Ongoing DeepSeek Distillation Attempts
OpenAI sent a memo to the U.S. House Select Committee on the CCP accusing DeepSeek of 'ongoing efforts to free-ride on the capabilities developed by OpenAI and other U.S. frontier labs.' It said accounts associated with DeepSeek employees had developed methods to circumvent OpenAI's access restrictions through obfuscated third-party routers, and that DeepSeek employees wrote code to obtain outputs from U.S. AI models for distillation programmatically. The memo explicitly described the activity it observed as ongoing attempts to distill OpenAI and other U.S. frontier models through 'new, obfuscated methods.' It sought government policy support on information sharing, router loopholes, adversary access to U.S. infrastructure and comparable standards among allies.
Anthropic Accuses DeepSeek of Industrial-Scale Distillation
Anthropic said DeepSeek, Moonshot AI and MiniMax ran industrial-scale campaigns to illicitly extract Claude's capabilities, generating over 16 million exchanges through approximately 24,000 fraudulent accounts in violation of its terms and regional access restrictions. Anthropic said DeepSeek's share was over 150,000 exchanges, focused on reasoning capabilities and on creating censorship-safe alternatives to politically sensitive queries.
U.S. Official Says DeepSeek Trained New Model on Banned Blackwell Chips
Reuters reported, citing a senior Trump administration official, that DeepSeek trained its upcoming model on Nvidia Blackwell chips despite the U.S. export ban, that the chips were believed to be in an Inner Mongolia data center, and that DeepSeek was expected to scrub technical signs of U.S. chip use before release. Nvidia declined to comment and DeepSeek did not respond.
DeepSeek Releases Open V4 Preview with One-Million-Token Context
DeepSeek released and open-sourced the V4 preview in Pro and Flash variants, both supporting a one-million-token context window, with the updated API available the same day.
DeepSeek Makes 75% V4-Pro Price Cut Permanent
DeepSeek said it would make permanent a 75% price cut on its flagship V4-Pro model, keeping prices at a quarter of their original level: 0.025 to 6 yuan per million tokens depending on usage type, down from 0.1 to 24 yuan.
DeepSeek Closes First Outside Funding Round Under Founder-Control Structure
Reuters, citing The Information, reported that DeepSeek raised more than 50 billion yuan ($7.40 billion) at a valuation above $50 billion in its first funding round. Investors put capital into a limited partnership managed by Liang Wenfeng rather than into DeepSeek, with a five-year lock-up and no voting rights; China's National Artificial Intelligence Industry Investment Fund was the exception, investing directly with voting rights. Reuters could not immediately verify the report.
U.S. Holds Off Adding DeepSeek to the Entity List
Reuters reported that DeepSeek, memory maker CXMT and more than 100 other companies had been approved by an interagency committee in 2025 for the Commerce Department's Entity List, but that the administration had held off publishing the additions to avoid escalating tensions with Beijing.
DeepSeek Announces Peak-Hour API Surcharge
DeepSeek told API subscribers it would double the price of V4 model access during peak hours (9am to noon and 2pm to 6pm Beijing time), raising V4-Pro output to 12 yuan per million tokens from the 6-yuan off-peak rate, citing resource distribution and service stability.
DeepSeek Reportedly Pauses Second Funding Round After Viral Founder Remarks
Bloomberg reported that DeepSeek told prospective investors it was suspending its second fundraising round, days after comments widely attributed to Liang Wenfeng about U.S.-China AI competition went viral, and that it might resume the process later.
DeepSeek Raises V4 API Prices More Than Fourfold at Peak Hours
New peak and off-peak pricing took effect on August 16, raising DeepSeek's V4 rates more than four times at peak, per a company post reported by Bloomberg. V4-Flash output rose from $0.28 to $1.32 per million tokens at peak (half that off-peak), and V4-Pro from $0.87 to $3.96. DeepSeek said the change would allocate resources more reasonably; rates remained below several rivals.
U.S. Agencies Name DeepSeek a Top Offender in Industrial-Scale Distillation
A joint CISA, NSA and FBI advisory said six Chinese AI companies, including DeepSeek, extracted billions of tokens through millions of requests from Anthropic, OpenAI, Google and xAI models since at least late 2024, using fraudulent accounts and proxies to evade restrictions. It identified DeepSeek and Moonshot AI as the top offenders and assessed that the scale indicated Chinese government awareness.
DeepSeek Hires CITIC Securities to Prepare STAR Market IPO
Reuters reported that DeepSeek hired CITIC Securities to prepare for an initial public offering on Shanghai's STAR Market and aimed to begin the process in 2026; timing, size and valuation were undecided.
Anthropic Says DeepSeek Silently Relayed Customers' Requests to Claude
Anthropic's September 2026 threat report said DeepSeek tagged users who reached its models through coding harnesses such as Claude Code and OpenCode and relayed selected users' requests to Claude Opus without informing them, using the exchanges to extract reasoning traces. Relayed data included a technology company's internal documentation and credentials tied to a Russian government database. Anthropic attributed more than 12.1 million exchanges to DeepSeek over 14 days in July 2026. DeepSeek did not respond publicly.
DeepSeek Releases V4.1-Flash and Lowers API Prices
DeepSeek released V4.1-Flash with native multimodal understanding, retired the previous Flash models, and reduced API prices. It had planned to route V4-Pro requests to V4.1-Flash from September 14, but said it would keep providing V4-Pro API service after that date in response to user demand, with billing unchanged.
China's CAC Reportedly Investigates DeepSeek Over Data Relayed to Anthropic
The Next Web, citing The Information, reported that the Cyberspace Administration of China summoned all seven labs named in Anthropic's report and was investigating DeepSeek and Moonshot AI over user data that may have reached Anthropic. Regulators had questioned staff; no penalty had been decided.
DeepSeek Starts Charging for Search and Fetch API Calls
DeepSeek opened its Search and Fetch APIs to all users at $0.75 per 1,000 calls for each endpoint, and said server-side Search and Fetch calls through DeepSeek Harness, Claude Code and Codex, free until then, would be billed from October 31, 2026.
DHS Orders DeepSeek Products Removed from Federal Civilian Systems
The Department of Homeland Security published notice of an order under the Federal Acquisition Supply Chain Security Act excluding DeepSeek and affiliate products from federal civilian procurements and requiring removal of DeepSeek products, and third-party products built using DeepSeek implementations, from civilian agency IT systems.
DeepSeek Nears Close of About 100 Billion Yuan Second Round
Reuters reported that DeepSeek had raised about 100 billion yuan ($14.98 billion) from more than 15 investors including CATL, Tencent, GigaDevice, Meituan and Geely, at a reported pre-money valuation of 500 billion yuan, doubling its original 50 billion yuan target. The round was expected to complete within days and details could still change.
Evidence (71 citations)
D1: User Value Erosion
D2: Business Customer Exploitation
D3: Shareholder Extraction
D4: Lock-in & Switching Costs
D5: Twiddling & Algorithmic Opacity
D6: Dark Patterns
D7: Advertising & Monetization Pressure
D8: Competitive Conduct
D9: Labor & Governance
D10: Regulatory & Legal Posture
Scoring Log (6 entries)
Checked 70 items + prose. 28 verified, 28 corrected (12 date-only), 14 re-sourced, 0 removed; 2 evidence items added (Tom's Hardware, DeepTech Asia). Invented: the House Select Committee 'high confidence' quote (report says 'highly likely'; chinaselectcommittee.house.gov DeepSeek Unmasked). Biggest fixes: timeline[18] was a misdated duplicate of the Feb 2026 OpenAI memo (now the Jan 2025 OpenAI/Sacks event); V2 price-war figures re-sourced to KrASIA; R1 wrongly given V3's $5.6M cost; 'all models MIT' corrected (pre-2025 weights used DeepSeek's own license); V3 capex is $1.6B per SemiAnalysis; Italy ordered a processing limitation, not a 'first GDPR emergency ban'; South Korea's 1.5M is an estimate; Cassidy-Rosen bill was May 2025; d3/d7 prose re-tensed (outside funding in 2026).
28→35. Since Feb 2026: first outside funding (>50B yuan, June 2026; ~100B yuan second round Oct 2026) and CITIC hired for STAR IPO; peak-hour API pricing raised V4 rates >4x at peak (Aug 16) after a June surcharge notice, alongside a permanent 75% V4-Pro cut (May) and V4.1-Flash cuts (Sep); paid Search/Fetch API (Oct); Anthropic alleged DeepSeek silently relayed customers' requests to Claude (Sep 10); CISA/NSA/FBI distillation advisory named DeepSeek (Sep 9); CAC probe reported (Sep 22); DHS FASCSA exclusion/removal order (Oct 20); Italy AGCM closed hallucination probe with commitments (Jan 2026); V4 open preview (Apr). D2 1→4 (event: Aug 2026 >4x peak price hike, June surcharge, paid Search/Fetch, alleged silent relay of customer requests to Claude). D3 2→3 (event: first outside funding rounds and IPO preparation). D6 2→3 (event: alleged bait-and-switch relay to Claude, partly offset by correction: Korea PIPC found keystroke patterns were never collected). D8 2→3 (event: U.S. joint advisory naming DeepSeek a top distillation offender; Anthropic Sept report). D9 4→5 (event: founder-control LP structure with state AI fund holding voting rights; State Dept military-support allegation). D1, D4, D5, D7, D10 unchanged. Eras: all four re-dated to inflection events: Research Lab Origins 2023-07-01→2023-07-17 (spin-off); Open-Source Breakout 2024-05-01→2024-05-06 (V2); Global Firestorm 2025-01-01→2025-01-20 (R1), summary expanded to absorb the 2025-early 2026 scrutiny formerly in 'Mounting Scrutiny'; final era 'Mounting Scrutiny' (dated to the 2026-02-20 assessment) re-dated to 2026-06-16 (first outside funding) and relabeled 'Outside Capital Turn'. Era scores re-derived: 5, 11, 28, 35. Alternatives unchanged (still accurate, no typed-in scores). Anthropic allegations are a competitor's claims and are framed as such.
Checked 11 trimmed candidates: 5 restored, 6 partly restored, 0 confirmed removed, 0 already present. timeline:update:0 PARTLY RESTORED — Exovera corporate profile: High-Flyer controlling shareholder and managing director, attributed to January 2025 profile; initial High-Flyer financing already present. timeline:update:9 RESTORED — Reuters (via Yahoo Tech): Less-than-an-hour remediation, attributed to Wiz CTO in Reuters. timeline:update:10 RESTORED — Freedom House, China Media Bulletin: Mid-sentence cutoff and erasure, attributed to Freedom House; corrected media descriptions retained. timeline:update:11 PARTLY RESTORED — Italy's Garante, January 30, 2025 order: Regulator findings on deficient disclosures, legal bases and China storage safeguards; absolute first and all claims withheld. timeline:update:13 PARTLY RESTORED — Promptfoo (study and published dataset): Tiananmen and Uyghur dataset topics only; corrected 1,360 sample and January 28 date retained. timeline:update:16 RESTORED — U.S. Congress / GovInfo, H.R. 1121 introduced text: 16 additional colleagues and 60-day OMB guidelines in the House bill; Cassidy-Rosen May timing retained. timeline:update:17 PARTLY RESTORED — Michael Bacarella (original Android static analysis): Android static-analysis code finding with execution caveat, separate from NowSecure; corrected anti-debugging attribution retained. timeline:update:19 PARTLY RESTORED — DigForCE Lab; Tom's Guide; Korea PIPC: Dated public privacy criticism and attributed general VPN warning; no proven biometric identification or causal backlash claim. timeline:update:20 RESTORED — Asia Times; Gasgoo: Dongfeng MHERO 917 cockpit integration with dated planned rollout. timeline:update:27 RESTORED — Bloomberg (via Bloomberg Law); CNBC: Reported intermediary-country route; Inner Mongolia already present as separate February 2026 event. timeline:update:28 PARTLY RESTORED — OpenAI, February 12, 2026 memo to House Select Committee: Original memo confirms ongoing distillation through new obfuscated methods; government policy requests, not a legislative demand. Read all original cited live pages; forced Wayback attempts failed and were not used as evidence. Used 20 web searches, no more than 3 per candidate; every restoration rests on fetched page text. No removals or score changes.