DeepSeek API Review 2026

What is DeepSeek API?

Half price after 3:30 pm IST. That is the quiet perk inside DeepSeek's rate card: on weekdays the API charges its full rate only during two windows (01:00 to 04:00 and 06:00 to 10:00 UTC), and every other hour costs 50% less. This DeepSeek API pricing and review page works through those rates, the two models currently on sale, the cache discount that makes repeated prompts almost free, and the data questions an Indian company should settle before sending its first request. Figures come from DeepSeek's pricing page, its rate-limit and caching guides, the privacy policy and the Open Platform terms as of October 2026.

Two models on sale: deepseek-flash and deepseek-v4-pro

DeepSeek, the Hangzhou lab founded in 2023 by Liang Wenfeng, sells access to its own models on a prepaid, pay-per-token basis. The price list is short. The name deepseek-flash points to DeepSeek-V4.1-Flash-1017, a build released on 17 October 2026, and it is the sensible default. deepseek-v4-pro runs DeepSeek-V4-Pro-0813 and costs about four times as much for fresh input. Both take a 1M-token context window and can write up to 384K tokens in one reply. Only Flash reads images.

If your code still calls deepseek-chat or deepseek-reasoner, it needs updating. Those two names served V3 and R1 in early 2025 and later pointed to V4-Flash; DeepSeek cut them off for good after 15:59 UTC on 24 July 2026. Older Flash aliases (deepseek-v4-flash and its vision-exp variant) still work, but the docs say they are now routed to V4.1-Flash-1017 and billed at the Flash rate.

DeepSeek API pricing per million tokens, peak and off-peak

All rates below are in US dollars per million tokens, from the official pricing page. Rupee figures are approximate conversions at ≈₹94.7 per dollar, because DeepSeek publishes no INR list.

Peak means 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, which is 6:30 to 9:30 am and 11:30 am to 3:30 pm in India. Weekends and Chinese public holidays are off-peak all day. This time-of-day split has been a standing part of the rate card since August 2026, not a limited offer. For an Indian team the move is obvious: queue bulk jobs such as catalogue tagging, call-transcript summaries or test-data generation for the evening, and the bill halves.

Two smaller meters exist. Search and Fetch, DeepSeek's in-house web search and page-retrieval APIs, cost $0.75 (approx. ₹71) per 1,000 successful requests each. Charges come out of a prepaid balance; if the account also holds a granted balance, that is used first. Under the Open Platform terms, refunds cover only unspent money, are reviewed on request and lose handling fees.

DeepSeek API cache hit pricing and how to earn it

Context caching is on for every account and needs no code change. When a new request begins with exactly the same prefix as an earlier one (a long system prompt, a product manual, the previous turns of a chat), those tokens are billed at the hit rate: $0.003 instead of $0.15 on Flash off-peak, a 98% cut. The usage object in each response reports prompt_cache_hit_tokens and prompt_cache_miss_tokens, so savings are visible call by call.

The caching guide is frank about limits. Matching works on whole prefix units, so editing one word near the top of a prompt breaks the hit for everything after it. Caches build within seconds and are cleared once unused, typically after a few hours to a few days, and DeepSeek calls the system best effort with no promise of a 100% hit rate. Put fixed instructions and reference text first and the user's question last.

DeepSeek R1 API reasoning now lives in thinking mode

R1, released in January 2025, is the model that made DeepSeek famous: open weights, visible step-by-step working, and maths and coding results that DeepSeek put close to OpenAI's o1. You can no longer call R1 itself on the hosted API. Reasoning is now a switch on both current models. Thinking mode is the default, and the quick-start example sets a thinking parameter along with a reasoning_effort value. Turn it off where a short, plain answer is enough. One catch for coding tools: fill-in-the-middle completion (still beta) works only with thinking switched off. R1's MIT-licensed weights remain on Hugging Face for anyone who wants to host it.

DeepSeek V3 vs GPT-4o API cost, redone with today's models

Older blog posts love this match-up, but both sides have moved. V3 is no longer sold on DeepSeek's API, and GPT-4o is being wound down: OpenAI's deprecations page shows its 2024-08-06 and 2024-11-20 snapshots shutting down on 14 January 2027, with GPT-6 Luna named as the replacement. The honest comparison today is deepseek-flash against GPT-6 Luna, with GPT-6.1 Sol as the step up.

The result surprises anyone who assumes DeepSeek wins every line. On OpenAI's price list, Luna charges $0.10 per million fresh input tokens and $0.50 for output, a little under Flash's off-peak $0.15 and $0.60. DeepSeek pulls ahead on repeated context, where a $0.003 cache hit beats Luna's $0.01 cached input, and on very long prompts, since OpenAI raises rates above 272K input tokens while DeepSeek lists no long-context surcharge. Peak hours tilt it back, because Flash doubles. OpenAI also halves rates through Batch and Flex; DeepSeek's pricing page lists no batch discount, so the off-peak window plays that role.

Take a support assistant that reads 30M input tokens a month (21M of them cache hits) and writes 3M. On Flash off-peak that costs about $3.21 (approx. ₹304); at peak rates, about $6.43. GPT-6 Luna comes to roughly $2.61 before OpenAI's cache-write charges, and GPT-6.1 Sol to about $50. deepseek-v4-pro lands near $12.34 off-peak. Between the two budget models the gap is small; against a mid-tier Western model it is large.

OpenAI compatible and Anthropic compatible out of the box

Most teams switch over by changing two settings. Point the OpenAI SDK at https://api.deepseek.com, or the Anthropic SDK at https://api.deepseek.com/anthropic, and send a DeepSeek key as a Bearer token. Both models support:

• JSON output through response_format set to json_object. The prompt must contain the word "json" plus an example, and the guide warns of occasional empty replies, so set a sensible max_tokens.
• Function calling (tool calls) in the familiar OpenAI style, which agent frameworks pick up without rewrites.
• The Responses API, chat prefix completion and FIM completion, the last two in beta.
• Agent setups documented for Claude Code, Codex and OpenCode.

What is missing matters just as much. There is no embeddings model, no speech model and no image generation, so a product that needs those will be paying a second vendor anyway.

Concurrency caps and why the DeepSeek API feels busy

Limits are counted in simultaneous open requests, not tokens per minute: 2,500 per account for deepseek-flash and 500 for V4-Pro, shared across all keys. Exceed them and you get HTTP 429. DeepSeek will raise the ceiling at no charge through a capacity request form, and expanded accounts can isolate traffic per user_id so one heavy customer does not starve the rest.

Slowness shows up differently. Under load a request may wait in a queue; the API holds the connection open by sending blank lines (non-streaming) or keep-alive comments (streaming), and drops it if inference has not started within 10 minutes. Hand-written HTTP parsers must skip those lines. Before blaming your own code during an outage, check status.deepseek.com.

Company data on servers governed by Chinese law

This is the part procurement will read first. DeepSeek's privacy policy (updated 10 February 2026) names Hangzhou DeepSeek Artificial Intelligence Co., Ltd. as controller and says personal data is collected, processed and stored in the People's Republic of China. It adds that personal data may be used to train and improve its models, with an opt-out right exercised by writing to privacy@deepseek.com. OpenAI states that API traffic is excluded from training by default; DeepSeek publishes no API-specific promise of that kind.

The Open Platform terms (effective 29 April 2026) hand you DeepSeek's rights in outputs and permit derivative products, even training other models on outputs. They also make you the controller of your own users' data, so DPDP Act duties sit with you. Disputes go to a court with jurisdiction over DeepSeek's Hangzhou office under mainland Chinese law, and the service comes as is, with no uptime SLA. For customer personal data, contracts under NDA or regulated financial records, the cleaner route is the open weights (MIT licence) on your own GPUs or in a cloud region you pick.

Signing up from India and getting a key

Nothing in DeepSeek's terms shuts out Indian developers; the restrictions cover sanctioned countries and parties under export rules. Government teams and their vendors should remember the February 2025 Finance Ministry advisory telling staff to keep AI tools such as DeepSeek and ChatGPT off office systems. Billing is a USD prepaid balance, and no standing free credit appears on the pricing page.

• Create an account on platform.deepseek.com, the developer console.
• Top up a small USD balance, then create a key in the console's API keys section and keep it server-side, since the terms make you liable for leaked keys.
• Set the base URL to https://api.deepseek.com and the model to deepseek-flash, and run a first test with thinking off.
• Log cache hit and miss tokens for a week before deciding whether V4-Pro is worth roughly four times the input price.

DeepSeek API pricing and review verdict: who should pay for it

It suits startups running high-volume, low-sensitivity text work (classification, summaries, code help, internal tools), products whose prompts repeat a large fixed context, and developers who like knowing the same model family can be self-hosted later. It suits SaaS firms holding Indian customers' personal data far less, along with enterprises that need an SLA, a signed DPA or certifications, and anyone who wants voice, images and embeddings on one bill.

Worth testing alongside it: OpenAI's GPT-6 Luna, now cheaper on fresh tokens; Anthropic's Claude Haiku 5.5 at the same $0.10 and $0.50; Google's Gemini API; or DeepSeek's own open weights on an Indian cloud region, which removes the China data question in exchange for running GPUs. Treat this DeepSeek API pricing and review as a dated snapshot, since model names shift often, and recheck the pricing page before fixing a budget.