DeepSeek's API Reference: What Developers Need to Know

I have everything I need. Now I’ll write the article.

---

DeepSeek’s API documentation doesn’t look like a threat. It’s a standard reference page — endpoints, authentication, rate limits, model names. There’s no breathless launch event, no CEO keynote, no carefully choreographed benchmark reveal. And that studied plainness is exactly the point. The most destabilizing thing DeepSeek has done to the AI industry isn’t a product announcement. It’s making a page of documentation cost more to read than most developers are comfortable admitting.

What the API Actually Offers

Strip away the noise and DeepSeek’s developer platform is a tight, focused offering. Three core models: V3.2 for general-purpose workloads, R2 for heavy reasoning tasks, OCR 2 for document extraction. Standard REST API. Context caching that cuts input costs by ~74% for repeated prompt prefixes — critical for any chatbot-shaped application sending the same system prompt thousands of times per day.

There’s no sprawling ecosystem attached to it. No storage products, no vector database upsells, no “AI platform” positioning that blurs the line between infrastructure and lock-in. DeepSeek’s API reference is short because the offering is deliberate. It does fewer things than OpenAI’s platform, and that’s the design choice.

The current pricing — a few cents per million tokens for V3.2, slightly more for R2 — sits at the kind of level where developers stop thinking about API costs as a variable to optimize and start treating them as noise. That’s a meaningful psychological shift when you’re building a product.

DeepSeek R2: The Reasoning Model That Shouldn’t Exist at This Price

R2 is the successor to R1 — the model that caused genuine distress in early 2025 when it matched o1’s benchmark performance at roughly 1/20th the cost. R2 extends that story. It’s a 671B Mixture-of-Experts architecture with 37B active parameters per inference call, trained for under $6 million, released with full weights under an MIT license, and available via the hosted API at pricing that makes sustained reasoning workloads — the kind that genuinely stretch o3 and Gemini 2.0 Flash Thinking budgets — suddenly affordable at startup scale.

91% on AIME 2026. Within three points of GPT-5 on that benchmark. 31% on FrontierMath. Tied with frontier Western models on GPQA Diamond. These aren’t cherry-picked numbers from a promotional release — they’re the benchmarks circulating among researchers who’ve run it themselves.

For developers building applications that require genuine reasoning depth — multi-step problem solving, mathematical verification, complex code generation — R2 is a model that demands a serious evaluation before you keep paying current o3 or Sonnet prices.

V4 Is About to Make This Conversation More Urgent

Everything above describes the current state. V4, targeting a late-April launch, resets the conversation.

The architecture: approximately 1 trillion parameters, MoE with ~37B active per token. Context window: 1 million tokens via what DeepSeek is calling Engram conditional memory. Native multimodal: text, image, and video in a single model. Three deployment tiers — Fast (Lite, latency-optimized), Expert (full model with deep reasoning), Vision (multimodal-first) — mirroring OpenAI’s reasoning vs. standard model split, but on economics that don’t compare.

Early API node tests from April have developers reporting 30% faster inference than V3.2, and 94% context recall accuracy at 128K tokens against V3.2’s 45%. SWE-bench performance at 81% would put V4 in direct competition with the strongest coding models currently on the market.

Pricing is expected around $0.30 per million input tokens, $0.50 per million output. That’s higher than the current V3.2 floor but still dramatically cheaper than Western frontier model pricing. The ratio between “best available Western model” and “competitive DeepSeek model” hasn’t been narrowing — it’s been staying stubbornly wide while the capability gap has been closing.

The Huawei Chip Angle Deserves Honest Treatment

Reuters reported in early April that V4 runs on Huawei Ascend 950PR chips. This isn’t a footnote.

The US export control policy of the last three years has operated on a theory: restrict Chinese access to advanced silicon, slow Chinese AI development. DeepSeek training and serving a trillion-parameter model on domestically produced hardware is a data point that challenges whether that theory is working as intended. The Ascend 950PR isn’t an H100 equivalent. DeepSeek’s algorithmic efficiency work — the MoE architecture, sparse activation, aggressive quantization — is compensating for hardware disadvantages in ways that weren’t widely predicted.

For developers making platform decisions, this has two implications worth thinking through clearly. First, a platform not dependent on Nvidia’s supply chain is structurally insulated from the GPU capacity crunches and pricing pressure that have hit Western providers. Second, routing production data through Chinese-operated infrastructure is a risk calculation that genuinely varies by use case. The right answer for a solo developer building a personal project is different from the right answer for a company processing medical records or financial transactions. Neither “it doesn’t matter” nor “it’s categorically off-limits” is a satisfying blanket position.

Compared to the Alternatives

The competitive picture is cleaner than vendor marketing tends to suggest.

OpenAI has responded to DeepSeek pricing pressure by releasing more capable models faster and defending the premise that frontier quality justifies frontier prices. There’s genuine substance to that argument — o3 and GPT-4o are excellent models. But “excellent at 20x the cost” is a position that erodes when production workloads compound over months.

Google’s Gemini Flash lineup is the most credible Western competitor on pricing. It’s fast, capable, and cheap enough that DeepSeek isn’t an automatic choice. The differentiation: Flash trades some reasoning depth for speed and cost, and V4’s million-token context window is longer than current Gemini Flash options. Gemini 2.5 Pro’s long-context positioning gets more complicated with V4 in the market.

Anthropic’s Claude — Haiku through Sonnet — sits at a higher price point with a stronger enterprise compliance story. SOC 2, data residency flexibility, and a reliability track record that regulated industries value. The audience is legitimately different. If compliance is load-bearing, Anthropic’s pricing premium is doing real work.

The honest summary: DeepSeek occupies a niche that didn’t previously exist cleanly — frontier-adjacent reasoning capability at commodity inference prices — and V4 extends that position rather than trading it away for more capability.

Verdict

DeepSeek’s API reference page is a quiet document making a loud argument. The argument is: the prices you’ve been paying were always too high, and you’ve been paying them because the alternatives were limited.

That argument was already strong. V4 makes it more uncomfortable. A trillion-parameter multimodal model with million-token context at $0.30 per million input tokens — if the benchmarks hold up against real-world workloads — puts serious pressure on anyone pricing their API above that floor and claiming the delta is justified by capability.

The geopolitical dimensions are real, and developers running production workloads should think through what “Chinese-operated infrastructure” actually means for their specific use case and company context. That’s not a reason to dismiss the platform — it’s a reason to make the decision consciously rather than defaulting either way.

The technical and economic case is straightforward: DeepSeek has built one of the best-value AI APIs available, V4 is about to make it better, and the gap between their pricing and Western frontier model pricing remains wide enough that not benchmarking your workloads against their current models is leaving money on the table. Late April is a reasonable deadline to do that math.