DeepSeek API pricing for V4 Flash and V4 Pro in 2026, with per-token cost tables compared to OpenAI GPT-5.5, GPT-5.4, and GPT-5.4 mini.
DeepSeek V4 Flash charges $0.14 per 1M input tokens and $0.28 per 1M output tokens. Those list prices already undercut every OpenAI model. But the real gap shows up on cache hits, where V4 Flash input drops to $0.0028 per 1M tokens, a 98% reduction that requires zero configuration on the developer's side.
For teams running AI agents or RAG pipelines with repeated system prompts, that automatic caching turns deepseek api pricing from "cheaper" into a structurally different cost model. This article breaks down the exact numbers for both DeepSeek and OpenAI, covers where each provider's pricing levers actually matter, and flags the legacy model deprecation deadline hitting later this month.
DeepSeek's API lineup in 2026 centers on two models.
V4 Flash is the workhorse. It supports a 1M token context window with up to 384K max output tokens, handles both thinking and non-thinking modes, and allows up to 2,500 concurrent requests. Pricing:
V4 Pro is the heavier reasoning model. Same thinking/non-thinking mode support, but with a concurrency limit of 500 simultaneous requests. Pricing:
Billing is purely token-based. No monthly subscription, no per-seat fees. You top up a balance, and token costs deduct from it directly.
The headline prices above tell one story. The cache-hit prices tell a different one.
When DeepSeek detects repeated content in your input tokens (shared system prompts, common prefixes, recurring context), it automatically serves those from cache at the reduced rate. V4 Flash drops from $0.14 to $0.0028 per 1M input tokens on cache hits. That is a 98% reduction, applied without any developer configuration.
OpenAI also offers prompt caching, but the cost structure differs. GPT-5.5 cached input runs $0.50 per 1M tokens. GPT-5.4 cached input is $0.25 per 1M tokens. GPT-5.4 mini drops to $0.075. Even OpenAI's cheapest cached rate ($0.075) is roughly 27 times more expensive than DeepSeek V4 Flash's cache-hit price ($0.0028).
This gap matters most in workloads where input tokens repeat heavily. Agent loops that prepend the same system prompt on every call. RAG pipelines that inject the same retrieval context across multiple queries. Chatbots maintaining long shared instructions. In those patterns, the majority of input tokens hit cache, and the effective cost per request collapses.
| Model | Input / 1M tokens | Cached Input / 1M tokens | Output / 1M tokens | Context Window |
|---|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.0028 | $0.28 | 1M |
| DeepSeek V4 Pro | $0.435 | $0.003625 | $0.87 | 1M |
| OpenAI GPT-5.4 mini | $0.75 | $0.075 | $4.50 | < 270K standard |
| OpenAI GPT-5.4 | $2.50 | $0.25 | $15.00 | < 270K standard |
| OpenAI GPT-5.5 | $5.00 | $0.50 | $30.00 | < 270K standard |
A few numbers worth sitting with. DeepSeek V4 Flash output at $0.28 per 1M tokens versus GPT-5.5 at $30.00 per 1M tokens: over 100x cheaper. Even against OpenAI's most affordable option, GPT-5.4 mini, V4 Flash output is roughly 16x less expensive.
On cached input, V4 Flash at $0.0028 versus GPT-5.4 mini at $0.075 is a 27x gap. Against GPT-5.5's cached rate of $0.50, the gap widens to 178x.
For a broader look at how Anthropic's pricing fits into this landscape, see our comparison of Anthropic and OpenAI API pricing.
Raw per-token rates matter less than how they interact with your actual usage pattern.
Agent loops with long system prompts. An agent that fires 50 LLM calls per task, each prepending a 4,000-token system prompt, sends 200,000 input tokens per task just on the prompt alone. With DeepSeek's automatic caching, most of those tokens hit cache after the first call. At $0.0028 per 1M cached input tokens, the system prompt cost across those 50 calls is negligible. On GPT-5.4 mini at $0.075 cached, the same pattern costs 27x more on input alone.
High-throughput RAG. If your retrieval pipeline injects a stable knowledge base prefix, the same cache dynamics apply. The larger the shared context relative to the unique query, the more DeepSeek's cache-hit pricing dominates.
General text generation without repetition. When inputs are unique on every call, cache hits are rare and list prices apply. V4 Flash at $0.14 input / $0.28 output is still cheaper than GPT-5.4 mini at $0.75 / $4.50, but the multiplier drops from 27x to about 5x on input and 16x on output.
Throughput constraints. V4 Flash supports 2,500 concurrent requests versus V4 Pro's 500. For burst workloads, Flash's higher concurrency ceiling matters independently of price. If your system peaks above 500 simultaneous requests, Pro becomes a bottleneck regardless of per-token cost. For routing strategies that handle these constraints automatically, see multi-provider LLM routing.
OpenAI's pricing is not limited to the per-token rates above. Two structural options shift the effective cost.
The Batch API offers 50% savings on both inputs and outputs for workloads that can tolerate asynchronous processing over a 24-hour window. Evaluation pipelines, bulk classification, document processing, anything that does not need a synchronous response can cut its OpenAI bill in half. GPT-5.4 mini on Batch API drops to $0.375 input / $2.25 output, which narrows the gap with DeepSeek V4 Flash on output (from 16x to 8x) while still remaining more expensive.
On the other side, data residency adds a 10% surcharge for teams that need regional processing guarantees. For compliance-driven buyers in regulated industries, this is a cost that does not have a DeepSeek equivalent.
Both of these levers are worth modeling into your AI FinOps planning alongside the base rates.
Teams still using the deepseek-chat or deepseek-reasoner model names have a deadline. Both will be deprecated on 2026-07-24 at 15:59 UTC.
After that date, deepseek-chat maps to V4 Flash in non-thinking mode, and deepseek-reasoner maps to V4 Flash in thinking mode. The mapping is already in place for compatibility, but the old names will stop working. If your codebase hardcodes either string, update it before the cutoff.
Running DeepSeek and OpenAI side by side creates a practical problem: two billing systems, two rate-limit regimes, two sets of API keys, and no unified view of what you are spending.
An AI gateway sits between your application and both providers, handling routing, budget enforcement, and caching through a single integration point. Instead of building provider-specific logic into your application layer, the gateway abstracts it. Route cache-heavy workloads to DeepSeek for the cost advantage. Route latency-sensitive or compliance-bound requests to OpenAI with Batch API or data residency. Set per-provider spend limits so a traffic spike on one provider does not blow your monthly budget.
If you are evaluating gateway options, our comparisons of LLM proxy architectures and LiteLLM alternatives cover the current landscape.
SHIM does not proxy DeepSeek. Use DeepSeek's native API or an integration that explicitly supports it, and keep any provider-selection or failover policy in your application.