DeepSeek API: Pricing, Benchmarks, and How to Access It

DeepSeek's API currently serves two models: deepseek-v4-flash and deepseek-v4-pro, both from the V4 family that replaced V3.2 this spring. The preview launched on April 24, 2026, the final V4-Flash landed on July 31, and V4-Pro reached general availability on August 13. Both models take 1M tokens of context, return up to 384K tokens of output, and publish open weights under the MIT license. New API pricing took effect on August 16, so most numbers you find in older posts are already wrong. Here is the current picture, verified against DeepSeek's own docs on August 18, 2026.

TL;DR: Yes, there is a public DeepSeek API at platform.deepseek.com. The current models are deepseek-v4-flash and deepseek-v4-pro, both with 1M context and 384K max output. DeepSeek's list pricing (since August 16, 2026) is peak/off-peak: V4-Flash at $0.44/$0.22 per 1M input and $1.32/$0.66 per 1M output, V4-Pro at $1.32/$0.66 in and $3.96/$1.98 out, with cache-hit input from $0.007 per 1M. Those rates are 3.5-4.7x the old flat prices, so any budget written before August needs a re-check. DeepSeek is not available on Unifically; the closest live alternatives here are Claude Fable 5, GPT 5.6 Sol, and Kimi K3.

Is there a DeepSeek API?

Yes. DeepSeek runs its own API platform at platform.deepseek.com, documented at api-docs.deepseek.com. You create an API key, top up a balance, and pay per token. There is no subscription.

The API speaks two formats from the same account:

• OpenAI-compatible: base URL https://api.deepseek.com, works with the standard OpenAI SDKs.
• Anthropic-compatible: base URL https://api.deepseek.com/anthropic, for tools built against the Anthropic API.

Both current models support JSON output, tool calls, and the OpenAI Responses API format, which DeepSeek added at the V4-Pro GA release with a one-click Codex setup. Thinking mode is on by default and can be switched off per request; since August 13 the thinking effort has three levels (low / high / max).

DeepSeek V4 release date and the current lineup

"DeepSeek 4" officially means the DeepSeek-V4 family. The release rolled out in stages during 2026, all dated in DeepSeek's change log:

So as of August 18, 2026, DeepSeek's latest model is DeepSeek-V4-Pro-0813, and the two live API models are:

• DeepSeek-V4-Pro: 1.6T total parameters, 49B active per token. The flagship, aimed at agent work and hard reasoning.
• DeepSeek-V4-Flash: 284B total, 13B active. The fast, cheap option, and by DeepSeek's own numbers surprisingly close to Pro on agent tasks.

Both use a sparse mixture-of-experts design with a new attention stack (token-wise compression plus DeepSeek Sparse Attention) that makes the 1M-token context window practical. The model card reports single-token inference cost at 27% of DeepSeek-V3.2's FLOPs in million-token contexts. Weights for both models are on Hugging Face under the MIT license, so self-hosting and fine-tuning are allowed, commercially included.

DeepSeek API pricing

These are DeepSeek's own list prices, billed by DeepSeek on platform.deepseek.com, effective since 16:00 UTC on August 16, 2026. They are not Unifically prices. The unusual part: rates change by time of day. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; every other hour is off-peak at half price.

Context caching is automatic. A cache hit on repeated input costs about 3% of the miss price, which matters a lot for agents that resend a growing transcript every turn. One worked example: a request sending 100K fresh input tokens and returning 10K tokens on V4-Flash off-peak costs about $0.029 ($0.022 input plus $0.0066 output). The same request at peak costs about $0.057.

Two operational limits worth knowing before you build on it: concurrency is capped at 2,500 parallel requests for V4-Flash and 500 for V4-Pro, and reasoning tokens bill as output, so thinking mode at max effort raises real output volume well past the visible answer.

If you schedule batch work, the off-peak window is a real discount: the same job costs half as much outside 01:00-10:00 UTC. Full details are on DeepSeek's pricing page.

How much did DeepSeek raise prices?

A lot. Until August 16, DeepSeek charged flat rates around the clock: $0.14 per 1M input and $0.28 per 1M output on V4-Flash, and $0.87 per 1M output on V4-Pro, per Reuters and Fortune. Against the new table, the output side looks like this:

Flash input moved the same way, from $0.14 flat to $0.44 peak and $0.22 off-peak. DeepSeek warned developers on August 6 that an increase was coming, and it had trialed time-of-day discounts in mid-July before making the two-tier structure official.

Two practical takeaways. First, the off-peak window now does the work the flat price used to do: for teams in US time zones, normal business hours land off-peak, so interactive traffic mostly sees the halved rates. Second, if your budget was built on the old $0.28 output rate, a peak-hour agent workload now costs about 4.7x what it did in July. Re-run the math before the invoice does it for you.

DeepSeek V4 benchmarks

DeepSeek published agent-focused numbers for both final checkpoints in its change log. Its own comparison shows why Flash is interesting: it holds most of Pro's agent capability at a third of the price.

V4-Pro also reports HLE at 42.7 without tools and 60.0 with tools. The Hugging Face model card adds, for Pro at max reasoning: 80.6 on SWE-bench Verified, 93.5 on LiveCodeBench, a 3206 Codeforces rating, and 83.5 on the 1M-token MRCR long-context test.

Read the fine print, though. DeepSeek notes the code-agent results were run on its own "DeepSeek Harness" in minimal mode at max effort, and that harness is not yet public. Until it is released and third parties reproduce the scores, treat the agent table as the vendor's numbers, not settled fact.

Independent crowd voting tells a more modest story. On Arena's Text leaderboard, pulled August 18, 2026, deepseek-v4-pro sits at 1458 Elo, rank #51 overall on 54,397 votes, and deepseek-v4-flash at 1435, rank #85. Claude Fable 5 leads the board at 1506, and Kimi K3 (max effort) holds #12 at 1489, which currently makes Kimi the stronger open-weights model by Arena's measure. V4 Pro's strength is agent benchmarks and price, not chat preference votes.

How to access the DeepSeek API

1. Sign up at platform.deepseek.com and top up a balance.
2. Create an API key under API Keys.
3. Point an OpenAI SDK at DeepSeek's base URL:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_DEEPSEEK_API_KEY",
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Summarize this changelog in three bullets."}],
)
print(response.choices[0].message.content)

Swap the model name to deepseek-v4-pro for the flagship. For Anthropic-format tools, use https://api.deepseek.com/anthropic as the base URL instead.

One migration note: the old deepseek-chat and deepseek-reasoner model names were retired on July 24, 2026. If an integration built before April still uses them, it now fails; update the model parameter to one of the V4 names.

Is DeepSeek on Unifically? Alternatives on one API

No. Unifically does not serve DeepSeek models today. If we add the DeepSeek V4 API to the catalog, this post will say so.

If you want a language model on the same key and balance you use for image, video, and audio generation, these are the closest live options:

• Claude Fable 5: #1 on Arena's Text board at 1506 Elo. The strongest general model you can call on Unifically, and the reference point V4-Pro gets measured against.
• Kimi K3: the closest analog to what DeepSeek offers, a huge open-weights MoE (2.8T parameters, 1M context) priced at $2.10 per 1M input and $10.50 per 1M output on Unifically. Our Kimi K3 review has the full benchmark picture.
• GPT 5.6 Sol: OpenAI's agent-focused model, a direct rival to V4-Pro on terminal and coding-agent work.

Live rates for every model are on the pricing page. Billing is pay-per-use with no subscription, and new accounts get $0.20 of free balance to test with.

FAQ

What is DeepSeek's latest model?

DeepSeek-V4-Pro-0813, released to general availability on August 13, 2026. It is the flagship of the DeepSeek 4 family; DeepSeek-V4-Flash-0731 is the smaller, cheaper sibling from July 31.

Is the DeepSeek API free?

No, it bills per token from a prepaid balance. The weights are a different story: both V4 models are downloadable from Hugging Face under the MIT license, so running them on your own hardware costs nothing in license fees.

What is the DeepSeek V4 context window?

1M tokens for both V4-Pro and V4-Flash, with a maximum output of 384K tokens. Context caching is automatic, and cached input bills at roughly 3% of the cache-miss rate.

Why did my deepseek-chat API calls stop working?

DeepSeek retired the deepseek-chat and deepseek-reasoner model names on July 24, 2026. Set the model parameter to deepseek-v4-flash (or deepseek-v4-pro) instead; the base URL is unchanged.

Does the DeepSeek V4 API cost more at certain hours?

Yes. Since August 16, 2026, DeepSeek bills peak rates during 01:00-04:00 and 06:00-10:00 UTC and half price at all other hours. Scheduling batch jobs off-peak halves the bill.

Did DeepSeek raise API prices in 2026?

Yes, sharply. On August 16, 2026, DeepSeek replaced its flat rates with peak/off-peak pricing. V4-Flash output went from $0.28 to $1.32 per 1M tokens at peak ($0.66 off-peak), and V4-Pro output from $0.87 to $3.96 at peak ($1.98 off-peak). The change ended the flat ultra-cheap pricing DeepSeek was known for.

Things that will change this post: independent replications once DeepSeek releases its evaluation harness, Arena ranks as votes accumulate on the 0813 checkpoint, any further V4 checkpoints or pricing revisions, and DeepSeek models arriving on Unifically. We will update it when they happen.