DeepSeek for Automation

Key takeaways

All facts are as of 12 October 2026.

• DeepSeek's current API models are V4.1-Flash (deepseek-flash, released 10 September 2026, natively multimodal) and V4-Pro-0813 (deepseek-v4-pro, GA 13 August 2026, 1.6T total and 49B active parameters). Both lines ship open weights under the MIT licence (every V4 model card and Hugging Face checkpoint tag read), with a 1M-token context and up to 384K output tokens.
• Off-peak list prices per 1M tokens are $0.15 input / $0.60 output for Flash and $0.66 / $1.98 for Pro; peak is double. Peak runs 06:30–09:30 and 11:30–15:30 IST on weekdays, so the US working day is entirely off-peak.
• On short prompts the small Western tiers (Claude Haiku 5.5 and GPT-6 Luna, both $0.10 / $0.50) now undercut Flash. DeepSeek still wins on cache hits, on long prompts (above 100K tokens against Haiku; against GPT-6 Luna only off-peak and above 272K), and at the mid tier, where V4-Pro off-peak costs about a quarter of Claude Sonnet 5.5 per task.
• An Anthropic-compatible endpoint needs only a base-URL change, but it silently maps any claude-* model name to a DeepSeek model.
• NIST's CAISI puts V4-Pro about 8 months behind the US frontier, close to GPT-5 overall, with its aggregate score between GPT-5.4 mini and Claude Opus 4.6. Its 2025 tests found R1-0528 agents 12 times likelier to be hijacked, and Cisco jailbroke R1 on 100% of 50 prompts. No public security test of V4 was found.
• The API is run by a China-based provider whose privacy policy (last updated 10 February 2026) says personal data is collected, processed and stored "in People's Republic of China" and used to train models, with an opt-out right; no API-specific retention period is stated. Several governments restricted DeepSeek on official devices or in app stores in early 2025, with 2026 status unknown, and no Indian government notice could be verified.
• The earlier V4-Flash (284B) self-hosts on 2–4 high-memory GPUs. V4.1-Flash is a 510 GB checkpoint and V4-Pro an 865 GB one that needs an 8-GPU datacentre node (Hugging Face file sizes). Without self-hosting, India-only processing exists only for older models: Amazon Bedrock runs DeepSeek V3.1 and V3.2 In-Region in Mumbai. Self-host to keep data in place, not to save money.

Scope, date and evidence labels

This report covers DeepSeek's models and API as of 12 October 2026; prices, terms, checkpoint sizes, hosting routes and benchmark positions were re-checked against primary sources on 12 October 2026. It is written for businesses in India, the US and elsewhere that are weighing DeepSeek for automation. The sources range from DeepSeek's changelog, pricing pages and model cards, arXiv papers and two NIST CAISI evaluations to security vendors, hardware guides and Wikipedia, so claims carry the labels below. Related documents in this set: landscape, top models, other models, hardware, costs and governance.

DeepSeek has shipped roughly every six weeks since mid-2024

DeepSeek's API changelog and Hugging Face repository creation dates give a dated release history (DeepSeek changelog; Hugging Face API, deepseek-ai). Releases came faster in 2026: V4 preview on 24 April, V4-Flash official on 31 July, V4-Pro GA on 13 August and V4.1-Flash on 10 September.

Sources for the table: DeepSeek changelog, DeepSeek news release list, Hugging Face API, Hugging Face API, V4-Flash-0731, Hugging Face API, V4-Pro-0813, vLLM day-0 post and DeepSeek V4.1-Flash card.

How DeepSeek handles model names matters more than how often it releases. When V4.1-Flash shipped, deepseek-v4-flash and deepseek-v4-flash-vision-exp were not switched off. They were re-routed to V4.1-Flash at Flash prices (DeepSeek pricing). Where the retired deepseek-chat and deepseek-reasoner names point now was not confirmed (DeepSeek changelog). A production workflow can therefore move to a new model generation with no code change, which silently invalidates any accuracy baseline you have measured or agreed. The timeline places these releases alongside those of other labs.

V4 buys cheap long context with sparse attention and small active parameter counts

V4 makes long context cheap. It is a mixture-of-experts design that activates a small share of its parameters for each token, and it uses compressed attention that shrinks the key-value (KV) cache. The glossary defines these terms.

Sources: arXiv 2412.19437, DeepSeek-V3; DeepSeek-V4-Pro card; DeepSeek-V4-Flash card; DeepSeek-V4.1-Flash card; vLLM, 7 Oct 2026.

The V4-Pro card claims that at 1M tokens V4-Pro needs 27% of V3.2's per-token inference FLOPs and 10% of its KV cache (vendor claim) (DeepSeek-V4-Pro card). vLLM puts V4.1-Flash's global KV cache at about 890 bytes per token in FP4 (vLLM, 7 Oct 2026). At that rate a full 1M-token context needs under 1 GB of cache (calc). The hardware document's 70B dense FP16 example needs about 0.3 MB per token, several hundred times more (Hardware). That is why DeepSeek can price 1M context so low, and why long-document work such as contract review, tender analysis or log triage suits it on cost.

There are two practical wrinkles. Vision is supported only by deepseek-flash, not deepseek-v4-pro (DeepSeek pricing), despite pre-launch reports that V4 would be multimodal (TechNode, 2 Mar 2026). And thinking is on by default on both API models (DeepSeek pricing), so switch it off for simple classification steps. DeepSeek's pages do not say how reasoning tokens bill (unresolved), whereas Anthropic, OpenAI and Google state that they bill as output (Anthropic extended thinking; OpenAI reasoning guide; Google Cloud pricing; Costs). One date is unresolved: the V4 report's arXiv ID (2606.19348) implies June 2026, but its abstract page says "Submitted on 26 Apr 2026" (arXiv 2606.19348).

MIT weights settle the licence question, but not the others

The V4 model cards read for this report (V4-Pro, V4-Flash and V4.1-Flash) carry the MIT licence. So do R1 and its distills, V3-0324, R1-0528, V3.1 and DeepSeek-OCR. OCR-2 is Apache-2.0, and Coder-V2-Lite is tagged "other" (Hugging Face API; DeepSeek-V4.1-Flash card). Hugging Face's metadata also tags the V4-Flash-0731 and V4-Pro-0813 checkpoints license:mit (read 11 October 2026) (Hugging Face API, V4-Flash-0731; Hugging Face API, V4-Pro-0813). The tag is metadata; the LICENSE file in each repository is the binding text.

For a business that wants to host the weights itself or build them into its own products, MIT puts DeepSeek ahead of some Chinese peers. Kimi K3 needs a separate agreement for model-as-a-service businesses above US$20M of revenue in 12 months (Kimi-K3 licence). MiniMax M3 requires a "Built with MiniMax M3" display plus a notice or authorisation for any commercial use (MiniMax-M3 licence). MIT weights carry no such thresholds, so fine-tunes, embedded products and hosted deployments are all open. Other models compares the licences in full.

The licence governs the weights, not the behaviour. The censorship and hijacking findings below were measured on the models themselves, so a self-hosted copy has the same weaknesses unless you add controls. Self-hosting fixes where the data goes. It does not fix how the model behaves.

Prices: off-peak halves the bill and cache hits cost almost nothing

DeepSeek's official table was re-read on 11 and 12 October 2026, and every price, the concurrency limits and the peak hours below matched. The page has no date stamp, only "Copyright 2026". It lists a 1M context and one 384K maximum-output row covering both models (DeepSeek pricing). INR uses ₹96 per USD (calc).

Prices are per 1M tokens. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Weekends and Chinese holidays are off-peak (DeepSeek pricing). Time-of-day pricing began at 16:00 UTC on 16 August 2026, and Flash prices were cut with the V4.1-Flash release on 10 September 2026. The pre-cut Flash price was not captured (DeepSeek changelog). The pricing page gives no end date for time-of-day pricing, and the 10 September release note does not call it permanent ("We will provide further notice should there be any changes"), so treat it as current pricing. In India the peak windows are 06:30–09:30 and 11:30–15:30 IST (calc), so batch jobs queued after 15:30 IST, overnight or at weekends pay half price. US business hours fall entirely in off-peak (calc). Concurrency limits apply per account, not per key, and DeepSeek raises them free on request (DeepSeek rate limits).

Sources conflict on the V4-Pro price. CAISI's April 2026 test used developer-listed prices of $1.74 uncached input, $0.0145 cached and $3.48 output (CAISI V4-Pro evaluation). The input figure is about 2.6 times today's off-peak price (calc). This report uses the official page because it is the primary and current source. The gap most likely reflects a reprice at the August GA, but the changelog does not state the old price. A $0.435 / $0.87 figure in secondary coverage is not a DeepSeek price: it is Xiaomi MiMo-V2.6-Pro's overseas list price (input cache miss $0.435, output $0.87) (Xiaomi MiMo pricing). OpenRouter lists V4-Pro-0813 at the off-peak rate but V4.1-Flash at the peak rate (OpenRouter models API), so check which tier any third-party price reflects.

DeepSeek's price lead has narrowed for small models and remains large in the middle of the market.

Sources: DeepSeek pricing; Claude platform pricing; OpenAI API pricing; Gemini API pricing.

On short prompts the cheapest Western tiers now undercut Flash on both uncached input and output. DeepSeek still leads in three places. The first is cache hits: $0.003 against Haiku 5.5's $0.01 matters for agent loops that re-send a long context every turn. The second is long prompts. Above 100K tokens Haiku 5.5 rises to $0.50 input (Claude platform pricing), while GPT-6 Luna keeps its $0.10 / $0.50 rate up to 272K input tokens, which undercuts Flash at any hour; above 272K the whole request reprices to $0.20 / $0.75, above Flash off-peak but below Flash peak (OpenAI API pricing). So the long-prompt lead holds at any hour against Haiku 5.5 above 100K tokens, but against GPT-6 Luna only off-peak and above 272K (calc). The third is the reasoning tier, where V4-Pro off-peak costs about a sixth of Opus 5.5's input price and a tenth of its output price (calc).

Worked cost example: a support agent and a coding agent

This example applies DeepSeek's four price tiers to two token profiles from this set's cost model (Costs). T2 is a five-turn support-resolution agent using 18,000 uncached input tokens, 12,000 cached and 1,500 output. T6 is a 40-turn code-change agent whose context grows from 20K to 120K, using 420,000 uncached, 2,380,000 cached and 140,000 output tokens, of which 80,000 are reasoning. Monthly volume is 10,000 resolutions plus 500 code changes. The DeepSeek rows are calc. The comparison rows come from the cost model, except Haiku 5.5 on T6. The cost model priced that run at $0.136 at Haiku's up-to-100K rate, but with context growing evenly from 20K to 120K the last 8 of the 40 turns exceed 100K tokens and bill at 5 times the rate, which brings the run to about $0.28 (calc, as in Claude; Claude platform pricing).

The example yields three results. Moving the same work from peak to off-peak halves the DeepSeek bill, saving about ₹10,900 a month on Flash and ₹43,600 on Pro at this volume (calc). V4-Pro off-peak costs about a quarter of Sonnet 5.5 and an eighth of Opus 5.5 for the identical profile (calc). And on short-prompt support work, Flash costs about a third more than Haiku 5.5, so DeepSeek is not automatically the cheapest option. On the long coding run the order flips off-peak: Haiku's price step above 100K tokens lifts it to about $0.28, against $0.154 for Flash (calc). The cost model's caveats apply in full. The profiles are assumptions, and they exclude retries, evaluations, observability and human review. The model's authors suggest multiplying by 1.2 to 2 for a production budget, itself an assumption (Costs). Leaving DeepSeek's default thinking mode on for simple steps also pushes output tokens above these profiles.

The Anthropic-compatible endpoint is useful, and it swaps models silently

DeepSeek offers an Anthropic Messages-style endpoint at https://api.deepseek.com/anthropic (DeepSeek Anthropic API guide). Since 13 August 2026 it has also supported the OpenAI Responses API natively (DeepSeek changelog). Agent code written against either SDK can switch with little change, including many of the harnesses covered in frameworks and tool stack. DeepSeek's guide describes the Anthropic-style endpoint as follows.

Source: DeepSeek Anthropic API guide, re-read on 11 October 2026.

The mapping is the governance problem. A Claude-named request sent to the DeepSeek base URL returns a working DeepSeek answer with no error, whether it came from a developer, a tool or a stray environment variable. DeepSeek's guide even presents the mapping as a way to get past the Claude Desktop app's model-name restrictions in developer mode by changing only the base URL and API key. One line of configuration can breach a contract that names the model provider. The ignored fields also change agent behaviour (inference): without is_error the model cannot see that a tool failed, sequential-tool assumptions break without disable_parallel_tool_use, Anthropic-style caching plans do not carry over, and MCP tools must be exposed as ordinary tool definitions. Pin explicit DeepSeek model IDs, keep base URLs on an allowlist, and log the provider and model on every call. The Claude document covers the native features this endpoint imitates.

Capability sits about eight months behind the US frontier

NIST's Center for AI Standards and Innovation (CAISI), which NIST's site now calls CAISSI (NIST hub), evaluated V4-Pro in April 2026 on cloud H200 and B200 GPUs with the developer's recommended settings, publishing on 1 May and updating on 2 May. It called V4-Pro "the most capable PRC AI model evaluated by CAISI to date", but found it about 8 months behind the US frontier. It also noted that V4 "scores better on DeepSeek's self-reported evaluations than on CAISI evaluations". DeepSeek's own reports put V4 near Claude Opus 4.6 and GPT-5.4 (vendor claim, as summarised by CAISI), while CAISI's results, which include non-public benchmarks, put it close to GPT-5, released about 8 months earlier (CAISI V4-Pro evaluation). The table below was re-read on the CAISI page on 11 October 2026.

Source: CAISI V4-Pro evaluation. CAISI imputed V4-Pro's CTF score from a subset of samples, and it notes that its SWE-Bench Verified scores tend to run lower than other evaluators' because of prompt, scaffold and token-budget differences.

V4-Pro is at or near the frontier on maths, where it beats Opus 4.6 on all three sets, and it trails by 5–7 points on the public SWE-Bench Verified set. It falls furthest behind on held-out engineering (PortBench), novel reasoning (ARC-AGI-2) and cyber tasks, which are the nearest proxies for unfamiliar real-world codebases. Its aggregate Elo sits just above GPT-5.4 mini and well below Opus 4.6. At the higher April prices, CAISI found V4 cheaper than GPT-5.4 mini on 5 of 7 benchmarks, ranging from 53% cheaper to 41% more expensive (CAISI V4-Pro evaluation). Critics note that it used only GPT-5.4 mini as the US reference and relied on private benchmarks (Decrypt).

Inflated claims circulate. A September 2026 aggregator snapshot credited V4-Pro-0813 with 96.4% on SWE-bench Verified, but the official leaderboard file tops out at 79.2% (December 2025 data) (SWE-bench leaderboard data). Artificial Analysis's own Intelligence Index (v4.3.2, live 12 October 2026) gives DeepSeek V4.1 Flash 39.5 and V4-Pro 36.0, against 57.6 for Claude Opus 5.5 (Max) and 46.3 for the best open-weight model, Xiaomi MiMo-V2.6-Pro (Artificial Analysis). Epoch's own SWE-bench Verified run puts V4-Pro at 77.6% (top: Claude Opus 4.7, 83.5%), close to CAISI's 74% and far below the 96.4% claim; Epoch's FrontierMath (Tiers 1-3) gives V4.1 Flash 67.4% (Epoch benchmark data). On LMArena's 8 October text snapshot V4-Pro ranks #62 (1458) (LMArena). This set's top ten for multi-step automation ranks DeepSeek V4.1-Flash seventh, mainly on cost and usage rather than capability (Top models). Treat DeepSeek-reported numbers as upper bounds, and evaluate on your own tasks before relying on any accuracy figure.

Self-hosting: V4-Flash fits on one node; V4-Pro needs a datacentre node

Neither V4 card lists hardware requirements. Both give quick-starts for vLLM, SGLang, Docker Model Runner and Transformers, link community quantisations for llama.cpp, Ollama and LM Studio, and recommend temperature 1.0 and top_p 1.0 (DeepSeek-V4-Pro card; DeepSeek-V4-Flash card). vLLM supported V4 on day 0 (vLLM day-0 post). The memory figures below come from secondary guides and vendor posts, except the V4-Pro and V4.1-Flash checkpoint sizes, which are Hugging Face file sums (864.7 GB and 510.3 GB; primary). Fit on any node is still arithmetic, so size from a real test.

Sources: BestLLMFor VRAM guide; Codersera VRAM guide; Yotta Labs V4-Flash guide (updated 15 Sep 2026); Lucebox, V4-Flash on Strix Halo; vLLM, 7 Oct 2026; vLLM day-0 post; Spheron deployment guide; ThunderCompute local guide.

Self-hosters now face a split. The API's V4.1-Flash is about three times larger on disk than the V4-Flash most self-hosters can afford (calc). Owner-run deployments on modest hardware are therefore likely to stay on V4-Flash-0731 while the API moves on. Downloads fit this: 4.44M for V4-Flash-0731 against 1.38M for V4.1-Flash (Hugging Face API). No source gives absolute tokens per second for V4 on H200 or B200 nodes. Unattributed guide claims, unchecked against vLLM release notes, call vLLM 0.9+ the only production-ready V4 runtime and report a vLLM compile crash on RTX PRO 6000 Blackwell cards around May 2026 (MindStudio guide; Codersera VRAM guide).

The economics favour the API. Indian commercial H100 rental runs about ₹217–362 per GPU-hour before 18% GST (July–August 2026 snapshot, third party) (ECorpIT). E2E Networks' H100 page listed ₹255.55 an hour on demand and ₹70 spot on 11 October 2026 (E2E Networks H100 pricing) (and "as low as ₹155.90" per GPU-hour on 3+ month commitments) (E2E Networks H100), and Lambda's self-serve H100 SXM (80 GB) costs $3.99 per GPU-hour on its largest H100 instance and $4.09–4.29 on smaller ones (Lambda pricing). Four 80 GB GPUs match the memory of the 4 x A100 80GB configuration that fits V4-Flash. Running them around the clock at the ₹217–362 rates costs about ₹6.3–10.6 lakh a month before GST. At off-peak Flash prices, the same money buys roughly 1.8 to 3.0 million T2-style resolutions (calc, model cost only). Buying instead of renting means two 96 GB RTX PRO 6000 Blackwell cards (NVIDIA RTX PRO 6000) at a launch price of $8,565, tracked street prices of about $11,100–11,700 (June–July 2026) and $16,000 on NVIDIA Marketplace by August 2026 (Tech Insider; GPU Poet, July 2026; secondary), plus an assumed $3,000–5,000 host. That totals about $20,100–37,000 (₹19.3–35.5 lakh), or about $25,200–28,400 (₹24.2–27.3 lakh) at tracked street prices (calc), before an assumed 0.25 FTE of operations time (Costs). For a developer pilot, a $3,149 (₹3.0 lakh) 128 GB Ryzen AI Max+ 395 mainboard (Framework) runs a heavily quantised V4-Flash at about 25 tok/s, which is enough for testing but not for production serving. Self-host DeepSeek to keep data in place, not to save money. Cost alone justifies it only at millions of tasks a month. The IndiaAI compute portal lists H100 SXM at ₹153.0 per GPU-hour before subsidy (GST extra; subsidy up to 40% subject to approval, so ₹91.80 at the maximum) for eligible users (IndiaAI price calculator). See hardware.

Hosted V4 runs outside India; India-hosted routes carry only older models

Sources: OpenRouter models API; Hugging Face API; AWS Bedrock DeepSeek page; Google Cloud models page; AWS regional availability; Azure region availability; E2E inference endpoints; Shakti Cloud AI endpoints; Yotta Labs V4-Flash guide.

Businesses that buy AI only through AWS or Google can therefore get DeepSeek V3.1, V3.2 or R1-0528, but not V4, on the pages read. Hosted V4 means DeepSeek's own API (data stored in China), OpenRouter, or Azure's Global deployment types (processing in any region). For data that must be processed in India, the managed options are older models: Bedrock's Mumbai In-Region V3.1 and V3.2, or the V3 and R1 endpoints of E2E and Yotta (vendor-reported). Check the live consoles before planning on V4 through a marketplace. See Asia-Pacific and India for the Indian GPU-cloud market.

The $5.6 million training figure covers one run and nothing else

Table 1 of the V3 paper prices the official V3 run at $2 per H800 GPU-hour. Pre-training took 2,664K GPU-hours ($5.328M), context extension 119K ($0.238M) and post-training 5K ($0.01M), for a total of 2,788K GPU-hours and $5.576M on a 2,048-GPU H800 cluster (arXiv 2412.19437 PDF). The paper says these costs "include only the official training of DeepSeek-V3" and then lists exclusions. The excerpt read here cut off partway through that list, so re-read the PDF before quoting it in full. By construction the figure is a rental-rate estimate for one final run. It leaves out earlier research and ablation runs, staff, data and the capital cost of the cluster.

R1 has its own, smaller bill: R1-Zero 101K H800-hours ($202K), SFT data creation 5K ($10K) and R1 41K ($82K), for a total of 147K hours and $294K at the same rate (arXiv 2501.12948 PDF). That total excludes V3-Base, on which R1 is built, and the A100 work used to prepare experiments. The peer-reviewed version appeared in Nature on 17 September 2025 (Nature). Even careful sources blur the two figures: Cisco wrote that R1 "was reportedly trained for approximately $6 million" (Cisco), mixing V3's number with R1's. The V4 report gives no training-cost figure at all (arXiv 2606.19348 PDF).

On chips, High-Flyer bought about 10,000 A100s before US restrictions (Wikipedia only) (Wikipedia, DeepSeek), and V3 and R1 were trained on H800s (arXiv 2412.19437 PDF; arXiv 2501.12948 PDF). The V4 report says the work runs "on both NVIDIA GPUs and HUAWEI Ascend NPUs platforms". That is a statement about kernels and platforms, not a disclosure of what V4 was trained on (arXiv 2606.19348 PDF). DeepSeek has published a DeepGEMM port for Ascend (GitHub, DeepGEMM-Ascend). A Bloomberg headline of 4 September 2026 reported plans to order more than 160K Ascend 950DT chips for a data centre in Mongolia (listing only) (HN search listing). The export-control picture keeps changing. Nvidia halted H20 production in August 2025 after Chinese directives, and China blocked H200 shipments days after the US approved them in January 2026. These come from Wikipedia citing press reports, and the current status is unresolved (Wikipedia, Nvidia). None of this changes what runs on your own GPUs, though it bears on how fast DeepSeek can add API capacity (inference). The 2025 launch moved markets: Nvidia fell about 17–18% on 27 January 2025, and the DeepSeek app topped the US iOS free chart (Wikipedia only) (Wikipedia, 2025 in AI).

Security audits found weak refusals and 12x agent-hijacking susceptibility

Three independent audits exist. None of them tested the security of the current V4 API or the V4 weights.

Sources: Cisco blog; NowSecure blog; CAISI report PDF, Sep 2025; CAISI V4-Pro evaluation.

For automation, the 12x hijacking figure matters most. Agents that read untrusted email, web pages or uploads and can also send messages, edit records or run code are exactly what hijacking attacks target. The figure applies to R1-0528 in 2025, and no public test shows whether V4 fixed the problem. Any DeepSeek agent with write permissions therefore needs your own prompt-injection tests before release. NIST's naming is inconsistent. Its hub page reads "Center for Advancing Innovation and Standards for Super Intelligence (CAISSI)", while the 2026 report uses "Center for AI Standards and Innovation (CAISI)" (NIST hub). This report follows the report's usage. See governance for agent-security controls.

Government restrictions are mostly 2025 device and app-store actions with unknown 2026 status

Every restriction below dates from 2025. Most come from Wikipedia articles that cite press reports rather than the decisions themselves. None was re-verified for 2026. All of them target government devices or distribution of the consumer app, not private use of the API.

Sources: Gottheimer press release; Introl; Wikipedia, DeepSeek (English); Wikipedia, DeepSeek (Italian); Wikipedia, DeepSeek (German).

The Berlin finding transfers most directly. It concerns moving personal data to China, which is the same question an Indian business faces under the Digital Personal Data Protection Act 2023 when it sends personal data to the DeepSeek API. Under the DPDP Rules, Rule 15 lets personal data be transferred abroad "subject to" requirements the Central Government specifies "by general or special order" about making it available to a foreign State or its agencies; the power to restrict transfers to notified countries is s.16(1) of the Act, by notification. Separately, Rule 13(4) lets the government require Significant Data Fiduciaries to keep personal and traffic data in specified categories in India, on a committee's recommendation. All of this applies from 13 May 2027 (G.S.R. 843(E); G.S.R. 846(E); Governance; India). No s.16 notification or Rule 15 order was found (MeitY's index was unreadable, so absence is not confirmed), and customer or partner contracts may be stricter than the law. US buyers may also weigh a report that in February 2026 Anthropic accused DeepSeek of using thousands of fraudulent accounts to generate millions of Claude conversations for training. The report comes from English Wikipedia, citing a New York Times article that was not read, and is unverified (Wikipedia, DeepSeek). See Americas and Europe for US and EU rules.

The company: a High-Flyer spin-off with outsized developer adoption

Almost all corporate facts come from Wikipedia. DeepSeek was founded on 17 July 2023 in Hangzhou as a spin-off of the AGI lab of the quant fund High-Flyer, which is its principal investor, and Liang Wenfeng is CEO (Wikipedia, DeepSeek). Liang was born in 1985, studied at Zhejiang University and co-founded High-Flyer. Wikipedia cites the Bloomberg Billionaires Index (July 2026) for his $36B net worth (Wikipedia, Liang Wenfeng). The company infobox gives 160 employees (2025). Ownership is unresolved: one article says Liang held 84% as of May 2024 through two shell companies, while the other says he holds 1% personally, with 99% in a High-Flyer partnership (Wikipedia, DeepSeek; Wikipedia, Liang Wenfeng). Wikipedia also reports, unconfirmed here, a May 2026 Series A of US$7B at $52B post-money (after April talks at a $10B valuation), July 2026 talks at $70B pre-money, and IPO preparations with a listing "as soon as 2027" or, per a cited Bloomberg headline, "as soon as this year" (Wikipedia, DeepSeek). See companies and people.

The usage data is firmer. In the week beginning 28 September 2026, DeepSeek led OpenRouter with 21.8% of text requests, ahead of Google (21.5%) and OpenAI (18.2%). V4.1-Flash was the top model by tokens over 30 days, at 101T. Anthropic had 2.5% of requests, yet Claude models headed the one OpenRouter per-task share-of-spend list captured (its "top models by task" section; the task was not recorded). OpenRouter's token leaderboards and request chart carry the caveat that "none of them measures users or spend" (OpenRouter rankings). Request share favours cheap, high-volume models, so it measures developer usage rather than enterprise share. It still shows DeepSeek has become a default for price-sensitive automation traffic. A "DeepSeek Harness" agent developer preview appeared on 13 August 2026 (listing only) (DeepSeek Harness), and DeepSeek Harness now heads the agent-integrations menu of DeepSeek's API docs (read 11 October 2026) (DeepSeek pricing). See use cases.

Risk matrix

The exposure ratings below are this report's own assessment for a business using DeepSeek in automation, based on the evidence cited above.

What this means for your business

Treat DeepSeek as a cost tier inside a multi-model architecture, never as the only engine, and choose the route for each request by the class of data it carries. Use the China-hosted API for public and non-sensitive work: its privacy policy stores data in China and allows training use unless you opt out. Use self-hosted MIT weights, or an India-hosted endpoint for an older model, when data must stay in India or in your own cloud. Use no DeepSeek where customer or partner policy, government rules, high-privilege agents or politically sensitive content are involved.

The OCR row draws on the DeepSeek-OCR repository (MIT, last push 27 Jan 2026). The other rows draw on the sources cited in earlier sections.

What it costs, and where it pays off

In the worked example, a five-turn support resolution costs $0.0036 on Flash and $0.0151 on V4-Pro off-peak. The saving is largest at the mid tier, where V4-Pro off-peak costs about a quarter of Sonnet 5.5 and an eighth of Opus 5.5, provided the CAISI capability gap is acceptable on your tasks. At the small tier, Western models match or beat DeepSeek on short-prompt prices, so the case rests on long context, cache-heavy agent loops and self-hosting. Indian batch jobs after 15:30 IST or at weekends pay half price, and US business hours are already off-peak.

Before you rely on it

• Gate by data class. The workflow, not the model, tags each request; only public or non-confidential requests reach api.deepseek.com, from an account that has exercised the training opt-out.
• Pin and test. Use explicit DeepSeek model IDs, never Claude names on the compatibility endpoint, and re-run a fixed evaluation set whenever DeepSeek re-routes a name, as it did on 10 September 2026. Run your own prompt-injection tests on V4 before giving any agent write tools, and check reasoning-token billing on a test bill.
• Record it. Name the provider and hosting location for each workflow, and record approval for any Chinese-origin model.

Questions to ask any vendor or implementation partner

Ask which data classes can reach the DeepSeek API and how that is enforced, which model IDs and base URLs are allowlisted and logged, what V4 injection tests were run, and which non-Chinese fallback is ready. See the playbook for implementation.

Open questions and unverified items

• India's position. No MeitY, CERT-In, RBI or other Indian notice on DeepSeek was verified. A February 2025 Finance Ministry advisory is sometimes mentioned, but it could not be verified and should not be relied on either way. Whether the Central Government will notify countries under DPDP Act s.16(1) or issue a Rule 15 order (both from 13 May 2027) is unknown; none was found, and MeitY's index was unreadable.
• 2026 status of every foreign restriction, and whether the US federal bill passed.
• API retention. The privacy policy (10 Feb 2026) states PRC storage and training use with an opt-out, but no API-specific retention period; the Open Platform terms (29 Apr 2026) let customers use inputs and outputs to train other models and state no retention period either.
• V4 security. No public jailbreak, agent-hijacking or censorship test of V4 or V4.1 was found.
• Reasoning-token billing. Not stated on the pricing page or the thinking-mode guide (re-checked 12 October 2026). (The 1M context and the 384K maximum output were confirmed on the pricing page on 11 October 2026, as one row covering both models.)
• Absolute throughput for V4-Pro or V4.1-Flash on H200, B200 or GB200. SGLang benchmarks were not reachable.
• V4 on hyperscalers and Indian clouds. No V4 found on Bedrock (Mumbai In-Region lists DeepSeek V3.1/V3.2), Google, E2E TIR or Yotta Shakti (both list V3/R1, vendor-reported). Azure AI Foundry's region tables list V4-Flash and V4-Pro as Global Standard deployments only (not India-bound); live deployment there was not tested. NVIDIA NIM is unverified. Bedrock Mumbai and Indian-cloud DeepSeek prices were not captured.
• India-hosted DeepSeek endpoints. E2E TIR's and Yotta Shakti's claims of serving from Indian data centres are vendor-reported; their data-handling terms and prices were not read.
• Legacy names. DeepSeek's changelog confirms deepseek-chat and deepseek-reasoner were to be discontinued on 24 July 2026, but no page read says where they route now (re-checked 12 October 2026).
• V4 training cost and hardware. Neither is disclosed; the Ascend role is confirmed only for inference platforms.
• Corporate facts. Team size, ownership, funding and IPO timing rest on Wikipedia.
• Math and Prover models. Not located; dates and specs unknown.
• DGX Station fit. The V4.1-Flash checkpoint is confirmed at 510.3 GB (Hugging Face file sum); against the DGX Station's 748 GB of coherent memory that leaves about 238 GB for KV cache and runtime on paper (calc). NVIDIA claims support for models "up to 1T parameters" (NVIDIA DGX Station), but no serving measurement exists.

Where sources disagreed, this report chose as follows.

Sources

1. DeepSeek API: Models and Pricing
2. DeepSeek API changelog
3. DeepSeek news release list
4. DeepSeek Anthropic API guide
5. DeepSeek Rate Limit and Isolation
6. Hugging Face: DeepSeek-V4-Pro model card
7. Hugging Face: DeepSeek-V4-Flash model card
8. Hugging Face: DeepSeek-V4.1-Flash model card
9. Hugging Face API: deepseek-ai models by downloads
10. arXiv 2412.19437: DeepSeek-V3 Technical Report (abstract)
11. arXiv 2412.19437: DeepSeek-V3 Technical Report (PDF)
12. arXiv 2501.12948: DeepSeek-R1 (PDF)
13. Nature: DeepSeek-R1 paper
14. arXiv 2606.19348: DeepSeek-V4 report (abstract)
15. arXiv 2606.19348: DeepSeek-V4 report (PDF)
16. vLLM blog: DeepSeek-V4 day-0 support, 24 Apr 2026
17. vLLM blog: DeepSeek-V4.1-Flash performance, 7 Oct 2026
18. NIST CAISI: Evaluation of DeepSeek V4 Pro, May 2026
19. NIST CAISI: Evaluation of DeepSeek AI Models, Sep 2025 (PDF)
20. NIST CAISSI hub page
21. Decrypt: US says China's best AI models lag behind
22. Cisco: Evaluating security risk in DeepSeek and other frontier reasoning models
23. NowSecure: Security and privacy flaws in DeepSeek iOS app
24. TechNode: DeepSeek plans V4 multimodal release, 2 Mar 2026
25. Kimi-K3 licence
26. MiniMax-M3 licence
27. Claude platform pricing
28. OpenAI API pricing
29. OpenRouter models API
30. OpenRouter rankings
31. SWE-bench leaderboard data
32. BestLLMFor: DeepSeek V4 self-hosted VRAM guide
33. Codersera: DeepSeek V4 VRAM and GPU requirements
34. Yotta Labs: DeepSeek V4-Flash hardware requirements
35. Lucebox: DeepSeek V4 Flash on Strix Halo
36. Spheron: Deploy DeepSeek V4 on GPU cloud
37. ThunderCompute: Deploy DeepSeek V4 locally
38. MindStudio: DeepSeek V4 with vLLM and SGLang
39. ECorpIT: India GPU cloud rental pricing 2026
40. E2E Networks: H100 pricing in India
41. Lambda pricing
42. NVIDIA RTX PRO 6000 Blackwell
43. Tech Insider: RTX PRO 6000 Blackwell price 2026
44. Framework Desktop mainboard, Ryzen AI Max+ 395
45. NVIDIA DGX Station
46. AWS: DeepSeek models in Amazon Bedrock
47. Google Cloud: models catalogue
48. E2E Networks: TIR inference endpoints
49. Yotta Shakti Cloud: AI endpoints
50. GitHub: DeepGEMM-Ascend
51. Hacker News search listing: Huawei Ascend 950
52. Wikipedia: Nvidia
53. Wikipedia: 2025 in artificial intelligence
54. Wikipedia: DeepSeek (English)
55. Wikipedia: DeepSeek (Italian)
56. Wikipedia: DeepSeek (German)
57. Wikipedia: Liang Wenfeng
58. Gottheimer and LaHood press release on DeepSeek government-device bans
59. Introl: DeepSeek government bans
60. DeepSeek Harness
61. GitHub API: DeepSeek-OCR repository
62. Hugging Face API: DeepSeek-V4-Flash-0731 metadata
63. Hugging Face API: DeepSeek-V4-Pro-0813 metadata
64. Google: Gemini API pricing
65. DeepSeek privacy policy (last updated 10 Feb 2026)
66. DeepSeek Open Platform terms of service (effective 29 Apr 2026)
67. Xiaomi MiMo: pay-as-you-go pricing
68. AWS: Amazon Bedrock regional availability by model
69. Microsoft: region availability for Foundry Models sold by Azure
70. e-Gazette: G.S.R. 843(E), DPDP Act commencement, 13 Nov 2025
71. e-Gazette: G.S.R. 846(E), DPDP Rules 2025, 13 Nov 2025
72. Artificial Analysis: model leaderboard (Intelligence Index v4.3.2, read 12 Oct 2026)
73. Epoch AI: benchmark data (zip)
74. LMArena: text leaderboard (8 Oct 2026 snapshot)
75. IndiaAI compute portal: price calculator
76. E2E Networks: NVIDIA H100
77. GPU Poet: RTX PRO 6000 Blackwell price, July 2026 (secondary)
78. Anthropic: extended thinking
79. OpenAI: reasoning models guide
80. Google Cloud: Gemini Enterprise Agent Platform pricing