TL;DR
DeepSeek is a strong pick for high-volume, non-sensitive enterprise work and a poor fit for regulated data sent to its hosted API. Its two API models in November 2026 are deepseek-flash for volume, images and agents, and deepseek-v4-pro for heavier text reasoning. Flash costs $0.15 per million input tokens and $0.60 per million output tokens off-peak, and peak hours bill at double. That is cheap, yet Claude Haiku 5.5 and GPT-6 Luna now bill less per uncached token at list price. NIST’s CAISI places V4 Pro about eight months behind the frontier and found an older model’s agents far easier to hijack. Use the hosted API for low-risk data, and run regulated work on self-hosted MIT weights behind a gateway.
Key Takeaways
Two API names still remain as of November 3, 2026: deepseek-flash (V4.1-Flash-1017) and deepseek-v4-pro (V4-Pro-0813), both with a 1M-token context window.
Off-peak prices are half of peak, and peak covers up to 35 of the 168 hours in a week, so moving batch jobs off-peak halves their bill.
A Flash cache hit costs $0.003 per million tokens against $0.15 for a miss, a 50x gap that especially rewards stable system prompts.
The vendor’s benchmark tables are self-reported, while NIST’s CAISI measured V4 Pro at roughly eight months behind leading U.S. models.
The privacy policy says DeepSeek processes and stores personal data in the People’s Republic of China, and the API terms name no other location, so keep regulated data off the hosted API.
MIT-licensed weights let teams self-host or fine-tune, but that moves security, evaluation and upgrade work onto your own team.
Watch on YouTube
AI Security for Enterprises: Kanerika’s CISO on Prompt Injection, ISO 27001 & SOC 2
Kanerika’s CISO explains how prompt injection reaches enterprise AI systems and which security controls and certifications matter when you deploy a model.
Cheap Tokens, Open Weights and a Trust Question
Picture a platform team pricing a support summarizer that reads 100 million input tokens a month. Priced on Flash at the off-peak rate, that input costs $15, less than a team lunch. The finance lead calls it a rounding error, so the pilot gets approved.
Then the legal team asks two questions. Where are the prompts stored, and who can read them?
The answer sits in a privacy policy rather than on the price card, and it changes which workloads can use the API at all. This guide covers both halves of that decision. It starts with the current models, the price mechanics and the independent test results. Then it compares three ways to deploy and ends with a 30-day plan.
What Is DeepSeek?
DeepSeek is a Chinese AI lab that builds large language models and sells API access to them. It also publishes the model weights, so anyone can download the same models and run them on their own hardware.
That mix explains the enterprise interest. A hosted API with low token prices suits quick pilots, while open weights give a route for data that must stay inside your walls. The lab now sits alongside the vendors in our ranking of the top LLMs , mostly on price and openness rather than on raw capability.
What sets it apart is the combination. Few vendors pair open weights under a permissive license with a fast release cadence. Fewer still add token prices near the low end of the market.
Open Weights Under MIT, Not Open Data
The MIT license covers the weights as well as the code for both current models, as the V4.1-Flash-1017 model card on Hugging Face shows. MIT is permissive, so you can run, modify, fine-tune and build products on the weights with few conditions.
Open weights are not open data, though. The training data is not published, so you cannot audit what the model learned or rebuild it from scratch. Keep that distinction in mind when you compare it with other open-source LLMs .
A Short History in Three Dates
Three dates show how fast the lineup moves. R1 reached the API on January 20, 2025, when it first drew wide enterprise attention. The V4 preview arrived on April 24, 2026, and V4.1-Flash then followed on September 10, 2026.
Less than two years separate the first and last of those dates. In between, the API names changed several times, so any integration written in early 2025 has already had to move.
DeepSeek Models in November 2026
Two model names are live on the API as of November 3, 2026, and each one maps to a dated model version that can change later. The mapping matters because the vendor can change the model behind a name without asking you to change any code.
For example, a call to the Flash name today reaches V4.1-Flash-1017, an update that did not exist in September. The table lists what the official pricing and models page states for each model, as well as the limits that shape capacity planning.
Detail deepseek-flash deepseek-v4-pro Model served DeepSeek-V4.1-Flash-1017 DeepSeek-V4-Pro-0813 Released V4.1-Flash on Sep 10, 2026; the 1017 update on Oct 17, 2026 General availability on Aug 13, 2026 Context window 1M tokens 1M tokens Maximum output 384K tokens 384K tokens Image input Yes No Concurrency limit 2,500 500 JSON output, tool calls, Responses API, Anthropic-format API Yes Yes Fill-in-the-middle (FIM) Non-thinking mode only Non-thinking mode only MIT-licensed weights Yes Yes
Both models speak the OpenAI-style and Anthropic-style request formats, so most client libraries work without a new SDK. That lowers switching cost in both directions, which is useful when you want a fallback model ready.
V4-Pro also exposes three thinking-effort levels: low, high and max. Higher effort spends more output tokens before the answer appears, so test whether max pays for itself on your tasks before making it the default.
API Model Names That Changed in 2026
The model names you see in older tutorials no longer map to what they once did. Here is the sequence from the official API change log .
April 24, 2026: the V4 preview is released.
July 24, 2026: deepseek-chat and deepseek-reasoner are discontinued, so older code must move.
July 31, 2026: V4-Flash enters public beta.
August 13, 2026: V4-Pro reaches general availability, with low, high and max thinking effort.
August 21, 2026: V4-Flash-Vision-Exp is added so Flash can take image input.
September 10, 2026: V4.1-Flash replaces V4 Flash, while V4 Pro continues after September 14.
October 17, 2026: the V4.1-Flash-1017 update ships behind deepseek-flash, once again without a name change.
Two legacy names still answer. Calls to deepseek-v4-flash and deepseek-v4-flash-vision-exp are still served by V4.1-Flash-1017 at the Flash price. That keeps old code running, and it also means your outputs can change on a date you did not choose.
Which DeepSeek Model to Start With
Start with Flash for most workloads. It reads images and carries five times the concurrency limit of Pro, while billing input at under a quarter of Pro’s rate.
Move to Pro for text-heavy reasoning where its larger model helps, such as long contracts or multi-step analysis. Test both on your own tasks, because vendor benchmarks rarely match your prompts. When image input drives the choice, our guide to multimodal AI explains how vision models are judged.
How DeepSeek Keeps Inference Cheap
The low prices rest on model designs that do less work per token. Both current models use a mixture-of-experts (MoE) design, and their model cards also disclose the main numbers.
Think of MoE as a large firm of specialists. Each question goes only to the few experts who know that topic, rather than to every employee. The firm stays big, but each answer stays cheap. The table summarizes what the cards for V4.1-Flash and V4-Pro report.
Model Total parameters Active per token Memory saving V4.1-Flash 552B backbone 8B in prefill, 16B in decode KV cache of 890 bytes per token, about a quarter of V4-Flash V4-Pro 1.6T 49B At 1M tokens, 10% of the KV cache and 27% of single-token FLOPs compared with V3.2
V4.1-Flash also changes the overall layout. It uses a causal encoder-decoder with a 20-layer encoder and a 20-layer decoder, and it also reads images natively. If those terms are new, our explainer on how LLM architectures differ walks through MoE as well as encoder-decoder designs.
The KV cache figure matters especially for long prompts. It is the memory a model holds for every token in context. A smaller cache fits more users on each GPU and makes a 1M-token window cheaper to serve.
None of this guarantees quality. It explains why the price card can sit low, and it tells self-hosting teams roughly how much memory to plan for.
DeepSeek API Pricing in November 2026
The API bills per million tokens, with separate prices for cached input, uncached input and output. Since August 16, 2026, every one of those prices also has a peak and an off-peak rate.
Price per 1M tokens Flash off-peak Flash peak V4-Pro off-peak V4-Pro peak Input, cache hit $0.003 $0.006 $0.022 $0.044 Input, cache miss $0.15 $0.30 $0.66 $1.32 Output $0.60 $1.20 $1.98 $3.96
These are list prices from the pricing page, checked on November 3, 2026. On every line the off-peak rate is exactly half the peak rate, so the clock decides the bill as much as the model does.
Read the card by starting with the cache-miss input price and the output price, because those two lines drive most bills. Then work out how much of your traffic can hit the cache and how much can run off-peak. The chart below previews the next section, pricing one fixed workload of 100 million input and 20 million output tokens across vendors.
When Peak Hours Apply
Peak hours run from 01:00 to 04:00 and from 06:00 to 10:00 UTC, Monday to Friday. Chinese public holidays are excluded, so those days bill at the off-peak rate.
That comes to up to 35 of the 168 hours in a week, about 21%, because holidays drop out. The windows land awkwardly for U.S. teams. In Eastern Standard Time they cover Sunday to Thursday evenings from 8 to 11 p.m. and weekday early mornings from 1 to 5 a.m.
Those are often the hours when nightly jobs run. Check your scheduler, because a batch window that feels off-hours locally can be peak time for the API.
Why Cache Hits Matter
A cache hit happens when the start of your prompt matches one the API has already seen. On Flash off-peak, a hit costs $0.003 per million tokens against $0.15 for a miss, a 50x gap.
Apply that to the opening example. Suppose 80% of those 100 million input tokens hit the cache. Monthly input then falls from $15 to $3.24, which is 80M tokens at $0.003 plus 20M at $0.15.
The design rule follows directly. Keep system prompts, tool definitions and long instructions stable at the front, then put the text that changes at the end.
Search and Fetch Billing
The vendor also sells Search and Fetch as per-request APIs, each at $0.75 per 1,000 successful requests. From 00:00 Beijing time on November 1, 2026, server-side tool calls made through DeepSeek Harness, Claude Code or Codex are also billed.
Agent workloads that search or fetch often therefore see these charges grow with every task. Track them as their own cost line, apart from tokens, so a busy agent does not hide inside a token budget.
DeepSeek vs GPT-6.1 and Claude 5.5 on Price per Million Tokens
A price comparison only means something when the workload is fixed. The table uses one: 100 million input tokens and 20 million output tokens a month, with no caching and prompts under 100K tokens.
Model Input per 1M Output per 1M Monthly bill Claude Haiku 5.5 $0.10 $0.50 $20 GPT-6 Luna $0.10 $0.50 $20 deepseek-flash, off-peak $0.15 $0.60 $27 deepseek-flash, peak $0.30 $1.20 $54 deepseek-v4-pro, off-peak $0.66 $1.98 $105.60 deepseek-v4-pro, peak $1.32 $3.96 $211.20 GPT-6.1 Terra $1 $5 $200 GPT-6.1 Sol $2 $10 $400 Claude Sonnet 5.5 $2 $10 $400 Claude Opus 5.5 $4 $20 $800
All figures are list prices, before caching or batch discounts, taken from the OpenAI API pricing page and the Claude pricing page . Long-context work also costs more, since GPT-6.1 Sol and Terra prompts over 272K tokens bill 2x input and 1.5x output.
The result surprises many teams. Flash is no longer the cheapest token, because Haiku 5.5 and GPT-6 Luna bill $20 on this workload against Flash’s $27 off-peak. The order flips on cached input. A Flash cache hit costs $0.003 per million tokens off-peak, against $0.01 for a Haiku 5.5 cache read or cached input on GPT-6 Luna. That puts Flash in the same price band as other small, fast models, which our guide to small language models also covers.
The vendor still wins clearly on price in the upper tier. V4-Pro off-peak costs $105.60 here, roughly half of GPT-6.1 Terra and far less than Opus 5.5 at $800.
Price is only half the comparison. For capability head-to-heads, read our Grok vs ChatGPT vs DeepSeek comparison as well as the ChatGPT vs Gemini vs Claude guide.
Kanerika Service
LLM Development Services
Kanerika helps teams choose models, model the token bill at real traffic, and build LLM applications that can switch between hosted and self-hosted routes.
Explore LLM Development How DeepSeek Performs: Vendor Benchmarks and Independent Testing
Two kinds of evidence exist on quality, and they should never be mixed because they answer different questions. One is the vendor’s own benchmark table, while the other is independent testing by a government lab.
Below are the scores the V4.1-Flash-1017 model card reports. The vendor chose the comparators, and newer GPT-6.1 and Claude 5.5 models are now available.
Benchmark (vendor-reported) V4.1-Flash-1017 V4.1-Flash GPT-5.6 Sol Opus 5 Terminal-Bench 4.0 42.3 31.2 39.9 51.8 DeepSWE v1.1 77.0 74.2 73.0 74.0 HLE without tools 42.6 36.8 44.5 56.3
Read the table for direction rather than rank. The 1017 update gains 11.1 points over V4.1-Flash on Terminal-Bench 4.0, yet it still trails Opus 5 on two of the three tests.
Vendor numbers still help with one job: tracking the gap between versions of the same model. A jump from 31.2 to 42.3 on one benchmark is a signal to re-run your own evaluation set. Even so, it does not prove the model now fits your task.
What Independent Testing Found
NIST’s Center for AI Standards and Innovation (CAISI) published its evaluation of V4 Pro in May 2026. It found that the model still lags the frontier by about eight months.
The lab’s own reports placed V4 near Opus 4.6 and GPT-5.4. CAISI, which also used non-public benchmarks, placed it near GPT-5 instead. It did find V4 more cost-efficient than GPT-5.4 mini on five of seven benchmarks.
An earlier CAISI evaluation from September 2025 tested R1-0528, then the lab’s most secure model. Agents built on it were 12 times more likely than evaluated U.S. frontier models to follow hijacking instructions. Under a common jailbreak, it answered 94% of overtly malicious requests, compared to 8% for U.S. reference models.
Those results come from an older model, so newer versions may behave differently. They still set the test plan: run hijacking and jailbreak prompts against any version before it touches tools or customer data.
Is DeepSeek Safe for Enterprise Use?
Safety for an enterprise buyer comes down to three questions. Where does your data go, how does the model behave under attack, and how mature is the vendor behind it?
The answers change with the access route, so treat safety as a deployment question, not a property of the model alone. Our LLM security guide covers the general controls, while this section covers what is specific to this vendor.
Where Your Data Goes
The vendor’s privacy policy states that personal data it collects is processed and stored in the People’s Republic of China. The data controller it names is Hangzhou DeepSeek.
For a regulated business, that single line often settles the question. Customer records, health data and anything under a data-residency rule should therefore stay off the hosted API. When residency is the core requirement, our guide to sovereign AI sets out the wider options.
The policy applies directly to the consumer app and website. Staff who paste work documents into the chat app create that exposure, even when your official integration avoids it. So permit or block the app explicitly in your acceptable-use policy.
Why DeepSeek Is Restricted in Some Places
Regulators moved quickly after R1 launched. On January 30, 2025, Italy’s data protection authority, the Garante, ordered an urgent limitation on processing Italian users’ data.
A day later, on January 31, 2025, Texas banned the app on the work devices of state employees and contractors. Check your own sector rules as well, since public-sector and defense contracts may carry device and vendor clauses.
Risk Depends on How You Access It
The same weights carry very different risk depending on the route. The consumer app and the hosted API both send prompts to the vendor’s servers. A managed cloud instead keeps traffic in your cloud account, while self-hosting keeps it on your own hardware.
Model behavior travels with the weights, though. The hijacking findings apply wherever the model runs, so agentic AI security controls matter on self-hosted deployments too.
Vendor maturity is the third risk. The pricing page lists no enterprise tier or service-level agreement, so plan your own fallback for any workload that cannot stop.
Three Ways to Run DeepSeek in the Enterprise
Enterprises reach these models through one of three routes. Each one trades data control against access to the newest model and the amount of work your team takes on.
Route Where prompts go Newest model available Ops effort Best for Direct API Vendor servers, with data stored in China Yes, both current models Low Public or low-risk internal data, such as quick pilots Managed cloud (Amazon Bedrock) Your cloud account No, since the page lists V3.1 and R1 Low AWS-first teams that need traffic to stay in-account Self-hosted MIT weights Your own infrastructure Yes, weights for both current models High Regulated data, air-gapped or sovereign setups
Amazon Bedrock runs the models as a fully managed, serverless option. As of November 3, 2026, its page lists V3.1 and R1 rather than V4.1-Flash or V4-Pro. The managed route therefore buys data control at the price of an older generation.
Teams already standardized on AWS should also compare Amazon Nova models on Bedrock , which run inside the same account boundary. For a broader view of keeping models in-house, see our guide to private LLMs .
A clear decision rule then follows from the table. Non-sensitive prompts may go to the hosted API, while regulated data runs on self-hosted weights or in-account cloud, and the gateway enforces which is which.
Nothing stops you from running two routes at once. A common split sends public text to Flash on the API and keeps internal records on self-hosted weights, while one gateway decides per request.
What Self-Hosting DeepSeek Actually Takes
Self-hosting gives the most control and the most work. Downloading the weights takes an afternoon, but serving them reliably at enterprise scale is a project with five parts.
Hardware: The V4-Pro-0813 model card shows the model served on a single node with four GB300 GPUs. Plan for that class of hardware for Pro, as well as spare capacity for failover. Flash has fewer active parameters, but its 552B backbone still spans several GPUs, so check its own card before sizing.Serving stack: The card’s vLLM example uses expert parallelism, an FP8 KV cache and DSpark speculative decoding, while SGLang runs it with –tp 4. Our comparisons of vLLM vs Ollama and SGLang vs vLLM help you pick the engine before you buy hardware.Security: Put the endpoint behind your identity provider, isolate it on the network, and then log every request and response.Evaluation: Build a task set from real prompts, then score it before go-live and after every change.Upgrades: Flash moved through three versions between July 31 and October 17, 2026, so budget time to test, swap and roll back weights.
Break-Even Rule for Self-Hosting
There is no universal break-even number, because GPU prices and staff costs vary too much between companies. The rule itself is simple.
Self-host when your data class forbids external APIs. Otherwise, self-host only when steady monthly token spend beats the fully loaded cost of the GPU node plus the staff who run it.
Run that sum with your own quotes. At the list prices above, a workload billing $27 a month on Flash is nowhere near the line. One spending tens of thousands of dollars a month, though, might be.
Fine-Tune or RAG?
Start with retrieval-augmented generation (RAG) for private knowledge, because it keeps facts in a store you control and can update daily. Fine-tune only to change format or behavior, such as a fixed output schema or a house tone.
The MIT license allows both on the open weights, so the choice is technical rather than legal. Our comparison of RAG vs fine-tuning covers the trade-offs in detail.
Kanerika Service
AI Governance Services
Kanerika sets up the classification, gateway, red-teaming and audit controls that let regulated teams run open-weight and hosted models with evidence.
Explore AI Governance How to Govern DeepSeek in Production
Governance turns a cheap model into one you can defend in an audit. The controls below apply to any model, but the release pace and test results here make each one pressing.
Gateway: Route every call through an LLM gateway so keys, quotas and policies live in one place.Data classification: Tag each workload by data class, and then block restricted classes from the hosted API at the gateway.Routing: Send work to Flash, Pro or a fallback model by task and data class, while keeping a fixed share of traffic on the fallback.Red-teaming: Run an LLM red-teaming set before every version change, including the hijacking patterns CAISI tested.Guardrails: Filter inputs and outputs with AI guardrails , especially for prompt injection, personal data and unsafe tool calls.Version pinning: Pin the model name in config, record the dated version it serves, and re-run evaluations when that version changes.Logging: Capture prompts, outputs, cost and latency with LLM observability tooling, then set retention by data class.Human approval: Require a person to approve any tool action that writes, sends or pays until red-team results justify more autonomy.
These controls belong inside a wider AI governance program with named owners, written policies and audit trails. Without a named owner, a gateway drifts within weeks.
The order matters as well. Classification and the gateway come first, because every later control depends on knowing which data each call carries.
DeepSeek Enterprise Use Cases: Where It Fits and Where It Does Not
The model family fits best where volume is high, data is low-risk and a wrong answer is cheap to catch. It fits worst when data is regulated or an error carries legal or financial weight.
Fits well Poor fit High-volume summarization of public or low-risk internal text Regulated customer, health or financial data on the hosted API Ticket, email and document classification Autonomous agents with write access before red-teaming Coding assistants on non-sensitive code Devices in public-sector bodies where the app is banned Internal RAG on self-hosted weights Workloads that need a contractual service-level agreement Batch jobs scheduled off-peak Any workload you cannot re-test when the model version changes
The pattern mirrors any enterprise LLM rollout. Start with contained, measurable tasks, and widen the scope only once the controls prove themselves.
Classification is usually the clearest early win. The label set is fixed, so accuracy is easy to measure and errors are cheap to catch by sampling. High volume also makes the low token price count.
Coding assistants need one extra check. Decide which repositories count as sensitive before anyone installs a plugin that sends code to an external API.
Checklist
AI Governance Checklist
Work through the policies, controls and owners an enterprise needs before routing production prompts to any LLM, hosted or self-hosted.
Get the Checklist → DeepSeek Limitations and Common Mistakes
Problems in production tend to come from assumptions rather than from the model itself. Seven mistakes are easy to make, but each one is avoidable.
Ignoring release churn: V4.1-Flash got a dated update 37 days after launch, so a prompt tuned in September may behave differently in October.Trusting legacy names: deepseek-v4-flash still answers, but V4.1-Flash-1017 now serves it, so outputs can shift without a code change.Assuming vision everywhere: Image input works on Flash only, while Pro accepts text.Sizing Pro like Flash: Pro’s concurrency limit is 500 against 2,500 for Flash, so plan for a 5x lower ceiling on bursty traffic.Expecting an enterprise contract: No enterprise tier or service-level agreement is published, so you own the fallback plan.Equating open weights with safety: MIT weights let you host and inspect the model, but not verify its training data or its behavior under attack.Forgetting peak pricing: Jobs that run during the peak windows pay double, which turns a $27 workload into $54.
Hallucination and tool access also need their own controls. Our guides on LLM hallucination and the Model Context Protocol cover grounding answers as well as scoping tool permissions.
How to Decide If DeepSeek Belongs in Your Stack
A 30-day evaluation answers the question with your own data instead of someone else’s benchmarks. Run it as four one-week steps, so each week ends with a clear output.
First week, define tasks and data classes: Pick three to five real tasks, collect at least 100 sample prompts for each, and mark which data classes every task touches.Second week, integrate through a gateway: Connect Flash, Pro and one comparison model behind the same gateway, so routing, logging and cost tracking match.Third week, test quality, cost and security: Score outputs against your rubric, record real token spend, and then run agent-hijack and jailbreak prompts.Final week, decide and pin versions: Choose a model per task, pin the model names, and write down what triggers a re-test.
Case Study
80% Ticket Auto-Resolution with LLM-Driven AI
A support desk now answers 80% of tickets automatically, with staffing costs down 70% and resolution time halved, using an LLM grounded in past tickets.
Read the Case Study → A Weighted Scorecard for the Final Call
Score the results on a weighted card so the decision is explicit and repeatable. An LLM evaluation framework gives you the test setup, while the weights below are a starting point to adjust.
Criterion Weight What to measure Task quality 30% Pass rate on your rubric, compared to a second model Security 25% Share of hijack and jailbreak prompts refused, and personal-data leakage Cost 20% Real monthly spend at your traffic, and the cache hit rate Data fit 15% Share of workloads whose data class allows the chosen route Operations 10% Latency, concurrency headroom and upgrade effort
A model that wins on cost but fails the security row should not go live for that task. Write that veto into the scorecard before the tests start, so nobody argues it away afterward.
How Kanerika Helps Enterprises Run Open-Weight and Hosted Models Safely
Kanerika treats model choice as a governance and cost problem first, so the approach is the same for DeepSeek, GPT-6.1 or Claude 5.5. The work runs in five stages, and each one produces something you can audit.
First, the team classifies workloads and data so it can decide which prompts may leave your boundary. Second, it picks an access route per class, such as the hosted API for non-sensitive volume and self-hosted MIT weights for regulated data. Third, it evaluates candidate models on your own tasks, together with red-team prompts modeled on the CAISI hijacking findings.
Fourth, every model sits behind a gateway with routing, version pinning and logging. Fifth, cost controls such as off-peak scheduling and cache-friendly prompts run alongside a re-test on every model update. Kanerika’s generative AI and AI governance teams run these stages. For business users, Karl also adds governed natural-language questions over enterprise data on top of the same controls.
What a Ticket Desk Shows About Model Choice
One engagement shows why the cost and deployment choice carries so much weight. The client, a B2B software company whose customers are small and midsize businesses in more than 40 countries, was losing support capacity to repetitive tickets. Kanerika turned its ticket archive into a knowledge store and routed new support requests through an LLM-based ticket resolver. Eight in ten tickets now get an automatic reply, staffing costs fell 70% and resolution time fell by half.
The case study does not name the model behind it, and nothing here suggests it was DeepSeek. Its value here is the shape of the workload. At ticket-desk volume, the per-token price, the route each data class takes and the fallback plan together decide both the bill and the risk.
Talk to Kanerika
Choosing Between Hosted and Self-Hosted Models?
Bring your workloads and data classes, and Kanerika will map each one to a route, a model and a test plan in a short working session.
Schedule a Demo → Wrapping Up
DeepSeek earns a place in an enterprise stack for high-volume, low-risk work, since Flash’s price and the open weights pay off there. It does not belong on the hosted API for regulated data, and every version needs red-teaming before it touches tools.
Prices and model names move often, so re-check the pricing page before each budget cycle. Should your enterprise use DeepSeek, then? Yes for contained, non-sensitive volume behind a gateway, and only on self-hosted weights, or the older models in an in-account cloud such as Bedrock, for anything regulated.
Frequently Asked Questions
What is DeepSeek?
DeepSeek is a Chinese AI lab that builds large language models and sells API access to them. It also publishes model weights under the MIT license. As of November 2026, its two API models are deepseek-flash, served by V4.1-Flash-1017, and deepseek-v4-pro, served by V4-Pro-0813. Both models support a context window of 1M tokens.
Is DeepSeek free to use?
The open weights are free to download under the MIT license, although running them needs your own GPUs. The API is paid per million tokens. In November 2026, deepseek-flash costs $0.15 for uncached input and $0.60 for output off-peak, and peak hours bill at double. Check the official site for current consumer app terms.
How much does the DeepSeek API cost in November 2026?
Off-peak, deepseek-flash costs $0.003 per million cached input tokens, $0.15 uncached and $0.60 for output. The deepseek-v4-pro model costs $0.022, $0.66 and $1.98 for the same lines. Peak rates are double and apply on weekdays outside Chinese public holidays from 1 to 4 a.m. and 6 to 10 a.m. UTC. Search and Fetch cost $0.75 per 1,000 requests.
Does DeepSeek support function calling and AI agents?
Yes. Both API models support tool calls and JSON output, plus the Responses API and an Anthropic-format API, so common agent frameworks connect without much work. Server-side tool calls through DeepSeek Harness, Claude Code or Codex are billed from November 1, 2026. Red-team agents first, because CAISI found an older model far easier to hijack.
Can you run DeepSeek through AWS or Azure instead of the DeepSeek API?
On AWS, yes. Amazon Bedrock lists DeepSeek-V3.1 and R1 as fully managed, serverless models, which keeps traffic in your own cloud account. Those models are older than V4.1-Flash and V4-Pro. For the newest models outside the vendor API, self-host the MIT weights in your own cloud. Check each cloud catalog for its current listings.
Does DeepSeek store data in China?
Yes, for its hosted services. The privacy policy states that personal data it collects is processed and stored in the People’s Republic of China, with Hangzhou DeepSeek named as the data controller. The API terms name no other location, so treat API prompts the same way. Self-hosted weights or Amazon Bedrock keep your prompts out of that path.
Is DeepSeek safe to use for business data?
It depends on the route you choose. The hosted API and app run on DeepSeek’s servers, and its privacy policy places personal data in China, so regulated data should stay off them. Self-hosted MIT weights keep data on your own hardware. NIST’s CAISI also found an older model easy to hijack, so red-team every version before it gets tool access.
Why is DeepSeek banned in some places?
Regulators acted over data protection and security concerns. Italy’s Garante ordered an urgent limitation on processing Italian users’ data on January 30, 2025. Texas banned the app on the work devices of state employees and contractors on January 31, 2025. The privacy policy says user data is processed and stored in the People’s Republic of China.
Is DeepSeek available in the US?
Yes. US developers can call the API and download the MIT-licensed weights, and Amazon Bedrock lists DeepSeek-V3.1 and R1 as managed models. Some public bodies restrict it, though. Texas, for example, bans the app on the work devices of state employees and contractors. Check your own sector and contract rules before any rollout.
Is DeepSeek better than ChatGPT?
On capability, independent testing says no. NIST’s CAISI placed V4 Pro about eight months behind the frontier, close to GPT-5. On price, the picture is mixed. V4-Pro costs far less than GPT-6.1 Sol, while GPT-6 Luna bills less than deepseek-flash per token. Test both on your own tasks before you choose one.
Is DeepSeek open source?
The model weights are open under the MIT license, which allows running, modifying, fine-tuning and commercial use. The training data is not published, so you cannot audit what the model learned or rebuild it from scratch. Open weights is the more precise term. Weights for V4.1-Flash-1017 and V4-Pro-0813 are available on Hugging Face.
What hardware do you need to self-host DeepSeek?
The V4-Pro-0813 model card shows the model served with vLLM on a single node with four GB300 GPUs. That setup uses expert parallelism, an FP8 KV cache and speculative decoding. SGLang can serve it with four-way tensor parallelism. Plan extra capacity for failover, and size Flash from its own model card before you buy hardware.
Which DeepSeek model should enterprises start with?
Start with deepseek-flash for most work. It accepts images, has a concurrency limit of 2,500 against 500 for Pro, and costs less per token. Move text-heavy reasoning, such as long contracts, to deepseek-v4-pro if your own tests show a gain. Run both against the same evaluation set before you commit to either model.
What changed between DeepSeek V4 and V4.1?
V4.1-Flash replaced V4 Flash on September 10, 2026. It uses a causal encoder-decoder design with a 552B backbone and a KV cache about a quarter the size of V4-Flash. It also reads images natively. The V4.1-Flash-1017 minor update followed on October 17, 2026, while the Pro model stays on its 0813 version.
