DeepSeek API Pricing vs Self-Hosted LLMs: Cost and Privacy (2026)

DeepSeek V4-Flash API pricing runs $0.14 per million input tokens and $0.28 per million output tokens (DeepSeek API docs), rates so low that the self-hosting math looks nothing like it did in our GPT-6 and Claude Opus 4.8 breakdowns. Those posts found self-hosting broke even somewhere between 16M and 108M tokens a day. Run the same math against DeepSeek's own API and the crossover moves into the hundreds of millions, sometimes billions, of tokens a day. Which raises the real question: if DeepSeek's API is already this cheap, why does anyone self-host DeepSeek at all? Cost isn't the answer. DeepSeek stores API data in China, trains on it by default, and offers no BAA or SOC 2 report, and no volume discount changes any of that.

DeepSeek API Pricing in 2026: V4-Flash, V4-Pro, and What Cache Hits Actually Save You

DeepSeek runs two models through its API today: V4-Flash at $0.14/$0.28 per million input/output tokens, and V4-Pro at $0.435/$0.87. Both support up to 1M tokens of context and 384K tokens of output, and both carry a steep automatic discount on any input DeepSeek's cache has already seen: $0.0028/M for V4-Flash, $0.003625/M for V4-Pro (DeepSeek API docs).

V4-Flash and V4-Pro Rates: Input, Output, Cache Hit/Miss

Source: DeepSeek API docs.

A $0.0028/M cache-hit rate against a $0.14/M cache miss is a 98% discount, and it applies automatically. DeepSeek's caching is disk-based; you don't flag which segments to cache the way you do with explicit cache-control breakpoints on some other APIs. Run the numbers on a real session: a 20,000-token system prompt reused across 50 calls costs $0.14 total if every call misses cache (1,000,000 tokens x $0.14/M). With the cache doing its job after the first call, that drops to about $0.0055, a 25x reduction, for doing nothing extra on your end.

The V4-Pro Discount Went Permanent, Not Just Extended

DeepSeek launched a 75% promotional discount on V4-Pro on April 24, 2026, cutting list price from $1.74/$3.48 per million input/output tokens down to $0.435/$0.87. Promotional pricing usually reverts once the window closes, and this one was scheduled to expire May 31. Instead, DeepSeek made the cut permanent on May 22, nine days ahead of the deadline, with "no rollback, no surprise expiry" (Apidog). Cache-hit input pricing fell by the same 75%, from $0.0145/M to $0.003625/M. Apidog's own math on the change: a team burning 50 million output tokens a day sees its monthly bill drop from roughly $5,200 to $1,300 (Apidog).

Legacy Model Deprecation: deepseek-chat and deepseek-reasoner Retire July 24, 2026

If your code still calls the deepseek-chat or deepseek-reasoner model names, that stops working on July 24, 2026 at 15:59 UTC (DeepSeek API docs), less than two weeks from this post's publish date. Both legacy names fold into V4-Flash: deepseek-chat becomes its non-thinking mode, deepseek-reasoner becomes its thinking mode. That mapping is worth sitting with for a second: it means DeepSeek's hosted "reasoning" endpoint now runs on V4-Flash's thinking mode, not the standalone 685B DeepSeek R2 checkpoint with its 10,000-40,000-token thinking chains. If your evals were tuned against R2's actual extended reasoning behavior, self-hosting R2 directly is the only way to get that exact model back.

The Self-Hosting Break-Even: Token Volume, Batch vs Real-Time, and Utilization Math

DeepSeek's own API runs roughly 20-60x cheaper per token than Claude Opus 4.8, so the self-hosting crossover isn't a modest step up from what we calculated for Opus, it's an order of magnitude higher. At an 80/20 input/output split, V4-Flash blends to $0.168/M and V4-Pro to $0.522/M uncached. Against the smallest realistic self-hosted cluster, that crossover lands somewhere between roughly 831M and 1.45B tokens a day for V4-Flash, and between 933M and 1.07B tokens a day for V4-Pro, depending on quantization tier and pricing mode.

Blended DeepSeek API Cost at 80/20 and 60/40 Input/Output Splits

At an 80/20 input/output split, the more common shape for RAG, agentic, and coding workloads:

V4-Flash blended (80/20) = (0.8 x $0.14) + (0.2 x $0.28) = $0.112 + $0.056 = $0.168/M tokens
V4-Pro blended (80/20)   = (0.8 x $0.435) + (0.2 x $0.87) = $0.348 + $0.174 = $0.522/M tokens

At a more output-heavy 60/40 split:

V4-Flash blended (60/40) = (0.6 x $0.14) + (0.4 x $0.28) = $0.084 + $0.112 = $0.196/M tokens
V4-Pro blended (60/40)   = (0.6 x $0.435) + (0.4 x $0.87) = $0.261 + $0.348 = $0.609/M tokens

Both ignore caching. If a meaningful share of your input hits DeepSeek's automatic cache, your real blended rate drops well below these numbers and the self-hosting crossover moves out further still.

One thing worth flagging up front: DeepSeek doesn't publish a separate batch-API discount the way Anthropic and OpenAI do. Synchronous and asynchronous requests bill at the same per-token rate, so batch vs real-time doesn't move your API cost the way it does in the Claude Opus comparison. What it does move is the self-hosted side: a rented cluster costs the same per hour whether it's running at 20% or 95% GPU utilization, so your real self-hosted cost per token depends entirely on how much of that fixed hourly spend you can actually fill with traffic. Batch workloads that can run around the clock get a much better effective rate out of the same cluster than bursty, low-utilization real-time traffic does.

Self-Hostable Alternatives: V4-Flash, V4, V3.2 Speciale, R2, and the GPU Cluster Each Actually Needs

All four are MoE models, so the full parameter count, not the active count, decides how much VRAM you need: the router can send any token to any expert, and every expert's weights have to be resident regardless of how few fire on a given forward pass. V4-Flash is the outlier that actually fits on a single node at a reasonable size. Its hybrid sparse attention mechanism, covered in detail here, is what lets it serve the full 1M-token context without KV cache blowing past what 4x H200 SXM5 can hold.

Notably absent from that table is V4-Pro itself. At roughly 1.6T total parameters, V4-Pro needs a minimum of 12x H200 SXM5 across two nodes, and the V4-Pro deployment guide walks through why: a single 8x H200 node tops out at 1,128 GB of VRAM, nearly 500 GB short of the FP8 weight footprint. Of the models covered here, V4-Pro is the one case where self-hosting is a genuinely harder lift than paying the API, and it's worth reading that guide before committing to it. For teams weighing DeepSeek's open weights against other open models entirely, DeepSeek vs Llama 4 vs Qwen3 has the cross-model cost and quality comparison.

GPU Cost Per Token: Cluster Cost, Not Single-GPU Cost

Self-hosted cost per token comes down to cluster price per hour divided by throughput. Live Spheron GPU pricing, fetched 11 Jul 2026:

Multiplying those rates by the minimum cluster each model needs gives the real daily infrastructure cost:

That cluster cost is fixed once provisioned, so it becomes relatively cheaper per token as daily volume climbs, and relatively more expensive as it falls below the cluster's throughput ceiling. Notice that H100 on-demand comes in cheaper than H100 spot in this snapshot, the opposite of what you'd normally expect. Spheron aggregates pricing across 5+ providers, and the cheapest bulk on-demand listing and the cheapest single-GPU spot listing don't always come from the same provider or region; H200 doesn't show the same inversion here. Always pull live pricing before locking in a tier, since this can and does shift. See GPU memory requirements for LLMs if you need to size VRAM for a quantization tier not covered here.

Daily Token Volume Crossover Table

The V4-Flash tier crosses over against a self-hosted V4-Flash cluster somewhere north of 800M tokens a day:

The V4-Pro tier, compared against the 8x H100 cluster that V4, V3.2 Speciale, and R2 all realistically need, crosses over in a tighter 933M-1.07B tokens-a-day band. On-demand pricing on that cluster is currently the cheaper of the two modes, so it crosses over first:

†Cluster cost is flat only up to that cluster's throughput ceiling; sustained volume well past the crossover may need a second cluster, which resets the comparison. Benchmark your own throughput before committing capacity at this scale.

Pricing fluctuates based on GPU availability. The prices above are based on 11 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

Put next to the other posts in this series, the gap is still wide. On production on-demand infrastructure, DeepSeek V4-Flash's crossover (1.45B tokens/day on 4x H100) sits about 66x higher than GPT-6's (22M tokens/day on H200 on-demand). The V4-Pro-tier crossover for V4, V3.2 Speciale, or R2 (933M tokens/day on 8x H100 on-demand) sits about 8.6x higher than Claude Opus 4.8's (108M tokens/day, also production FP8 on-demand). DeepSeek's API pricing is simply that aggressive: at these rates, pure cost rarely justifies self-hosting unless you're running a genuinely large production workload, and even then, prompt caching on the API side will claw back a lot of that gap before you provision a single GPU.

Privacy, Data Residency, and Compliance: Why Some Teams Can't Use the API at Any Price

No token volume changes this part of the decision. DeepSeek stores API prompts on servers in China, where Article 7 of China's National Intelligence Law compels organizations to cooperate with state intelligence work. DeepSeek also trains on API data by default, publishes no SOC 2 report we could find, and offers no BAA. For regulated or IP-sensitive workloads, that's a self-hosting decision, not a cost one.

Where DeepSeek API Data Actually Goes: Servers in China, Article 7 of the National Intelligence Law

DeepSeek's privacy policy is direct about this: "To provide you with our services, we directly collect, process and store your Personal Data in People's Republic of China" (DeepSeek privacy policy). That includes prompts and any files you upload through the API.

Article 7 of China's 2017 National Intelligence Law states that "all organizations and citizens shall support, assist, and cooperate with national intelligence efforts in accordance with law, and shall protect national intelligence work secrets they are aware of" (Wikipedia, National Intelligence Law of the People's Republic of China). This isn't a hypothetical concern raised only by critics: South Korea's Personal Information Protection Commission separately found that DeepSeek transferred prompts to Beijing Volcano Engine Technology and other Chinese companies without user consent (WitnessAI). Once prompt data is stored under that legal framework, no contractual clause from DeepSeek can override the underlying law.

No BAA, No SOC 2, and Training on API Data by Default

Two more gaps compound the residency issue. First, DeepSeek's terms allow training on user-submitted API data by default. A comparison from Tokenmix lays it out plainly: DeepSeek is "allowed by default," while OpenAI, Anthropic, and Google all exclude API data from training by default (Tokenmix). DeepSeek's own privacy policy confirms it processes data "to train and improve our technology, such as our machine learning models and algorithms," and grants an opt-out right specifically to users in the European Region; it doesn't clearly extend the same opt-out elsewhere (DeepSeek privacy policy). Once a prompt is baked into model weights through training, you can't selectively remove it later.

Second, there's no BAA. A review of DeepSeek's Open Platform Terms of Service found that "neither the DeepSeek Open Platform Terms of Service nor the privacy policy mentions HIPAA, PHI, or a Business Associate Agreement, and no BAA request channel was found" (AI Provider Trust Registry). We found no public SOC 2 report or trust center for DeepSeek's first-party API either, unlike Anthropic, OpenAI, and Google, which all publish one. Without a signed BAA, an organization legally cannot share protected health information with DeepSeek's API regardless of any other safeguard in place.

Enterprises that need DeepSeek's models under real compliance guarantees generally don't use DeepSeek's own API at all. Microsoft, AWS, and Google have all added DeepSeek to their managed model catalogs, and each inherits that cloud provider's own framework rather than DeepSeek's: Amazon Bedrock hosts DeepSeek-R1 with the same data-isolation guarantee it applies to every model on the platform, stating "your data is not shared with model providers, and is not used to improve the models" (AWS); Azure AI Foundry lists DeepSeek-R1 as a "Direct from Azure" offering backed by Microsoft's own enterprise infrastructure and support (Microsoft Azure AI Foundry); and AWS Bedrock, Azure AI Foundry, and Google Vertex AI all now offer DeepSeek models from EU regions specifically so European customers can use them under GDPR, with data kept in-region (innFactory). That's a real option worth knowing about if you need DeepSeek's outputs but not DeepSeek's own data terms, though it comes at whatever markup the hyperscaler charges over DeepSeek's own API pricing.

When Self-Hosting Is the Only Option, Regardless of Volume

Three categories of team should self-host DeepSeek regardless of what the token-volume math says: anyone under HIPAA who needs a real BAA and can't get one from DeepSeek or wait on a hyperscaler procurement cycle, anyone with contractual or regulatory data-residency requirements that specifically exclude China as a storage jurisdiction, and anyone handling IP-sensitive prompts (proprietary source code, unpublished research, M&A material) where "the vendor might train on this" is disqualifying on its own. For EU-based teams specifically, the EU AI Act compliance guide for GPU cloud covers the broader data governance obligations that sit on top of this. On Spheron, self-hosted instances run with SSH root access, no shared GPU tenancy, and no prompt or completion logging on the provider side, which is the baseline you want before treating self-hosting as a compliance shortcut rather than just a cost lever.

Self-Hosting DeepSeek on Spheron: GPU Sizing, VRAM, and the Real Hourly Cost

Direct answer: if you've decided to self-host on privacy grounds, sizing comes down to total parameter count at your chosen precision, not active parameters. V4-Flash is the only DeepSeek model that fits comfortably on a 4-GPU node; V4, V3.2 Speciale, and R2 all need an 8-GPU node at minimum; V4-Pro needs a two-node, 12-GPU cluster. Pick the smallest model that clears your quality bar rather than defaulting to the largest.

VRAM and Minimum Cluster Size by Model (V4-Flash, V4, V3.2 Speciale, R2)

At FP8 (1 byte per parameter), weight VRAM is roughly the total parameter count in gigabytes, plus KV cache headroom on top:

• V4-Flash (284B total): ~284 GB weights. 4x H100 SXM5 (320 GB) covers short context; 4x H200 SXM5 (564 GB) is needed for the full 1M-token window.
• V4 (~1T total): ~500 GB weights at FP8. 8x H100 SXM5 (640 GB) is the floor.
• V3.2 Speciale (685B total): ~640 GB weights. 8x H100 80GB (640 GB) is the minimum viable, tight on KV cache; 8x H200 (1.13 TB) is the production recommendation for full 128K context.
• R2 (~685B total, provisional): ~685 GB weights. 8x H100 SXM5 (640 GB) is the stated floor; 4x H200 SXM5 gives more KV headroom for the long thinking chains R2's reasoning mode generates.

Getting the parameter count right matters more than getting the active-parameter count right. All four models route through a large pool of experts per token, and because the router can select any expert on any forward pass, every expert's weights have to sit in VRAM whether or not they fire on a given request. That's why a 13B-active model like V4-Flash still needs a real multi-GPU cluster, not a single card.

Live Spheron GPU Pricing and Daily Cluster Cost

The cluster cost tables earlier in this post use Spheron's live H100 SXM5 and H200 SXM5 rates, aggregated across 5+ providers so you're comparing real market pricing rather than list rates from a single data center. To provision a cluster, browse the H100 GPU rental or H200 GPU rental catalog, pick spot for cost-sensitive batch workloads or on-demand for anything customer-facing, and follow Spheron's SSH connection guide to get root access to the instance. None of these clusters need anything beyond standard vLLM tensor and expert parallelism to serve; the individual deployment guides linked throughout this post cover the exact flags for each model.

Decision Framework: API vs Self-Host

For the two other posts in this API-vs-self-host series, see GPT-6 vs self-hosted LLMs and Claude Opus 4.8 vs self-hosted LLMs. If you're migrating a production workload off a closed API entirely rather than just running the numbers, migrating from the OpenAI API to self-hosted LLMs covers the operational side, and AI Inference Cost Economics 2026 covers the FinOps view beyond a single breakeven table. Spheron's docs cover instance provisioning in more depth if you're standing up your first cluster.

If your DeepSeek usage is climbing toward hundreds of millions of tokens a day, or your data can't legally sit on a server in China, it's worth benchmarking V4-Flash or R2 against your own eval set on a rented H100 or H200 cluster before your next contract renewal.

H100 GPU rental → | H200 on Spheron → | Check current GPU pricing →