DeepSeek Pricing Changelog 2026

DeepSeek's API changed in four ways through mid-2026: deepseek-chat and deepseek-reasoner now resolve to deepseek-v4-flash, deepseek-r1 stopped being callable, the cache-hit price was cut 10x on 2026-04-26, and the V4-Pro 75%-off rate that was supposed to expire on 2026-05-31 is still the live price. If you are pricing DeepSeek off a 2025 article, almost every number is stale.

This is a changelog, not a rate card. We called api.deepseek.com directly, recorded the raw usage objects, and matched them against the official docs so you can see what actually changed versus what older blog posts still claim. The short version: the model lineup collapsed to two names, the cheap one got cheaper, and the discount everyone expected to vanish stayed put.

What changed in the DeepSeek API (verified 2026-06)

According to DeepSeek API Docs, the official endpoint now exposes exactly two real model names, deepseek-v4-flash and deepseek-v4-pro, and prices them per million tokens. Everything else you may have called is now an alias or an error. Here is the timeline of changes that move your bill, each one confirmed against a live call rather than copied from a press release.

Change	What it was before	What it is now (2026-06-25)	How we confirmed it
Alias resolution	deepseek-chat billed as legacy V3 ($0.27/$1.10)	deepseek-chat and deepseek-reasoner both serve deepseek-v4-flash ($0.14/$0.28)	Live call returned served_model deepseek-v4-flash
deepseek-r1	Separate purchasable model ($0.55/$2.19)	Not callable; HTTP 400, folded into V4	Live call returned a 400 error
Cache-hit price	Launch-day cache rate	Cut to 1/10 on 2026-04-26, now $0.0028/M	Official docs plus live cache test
V4-Pro pricing	Nominal list $1.74/$3.48 (never took effect)	$0.435/$0.87, the standing live price; no reversion happened	Re-verified against official docs 2026-06-25

The practical upshot is that the model name you type no longer tells you the model you get, and two of the four old price points buyers still quote are now fiction.

Change 1: deepseek-chat and deepseek-reasoner became aliases

The biggest silent change is alias resolution. When we sent a request to deepseek-chat, the API answered with served_model: deepseek-v4-flash. Same story for deepseek-reasoner. The names survived for backward compatibility, but they no longer map to the older V3-generation models people still price against. A model alias is a legacy name that the API silently redirects to a current model, so the rate you pay is the current model's rate, not the name's historical one.

The legacy V3 figure of $0.27 input and $1.10 output that some 2025 articles still quote for deepseek-chat is stale. Both aliases now bill at V4-Flash rates of $0.14 input and $0.28 output per million tokens. If you budgeted against the old number, you are over-estimating input by roughly half and output by nearly 4x, so this correction works in your favor. For the line-item breakdown, see our input vs output pricing note.

Change 2: deepseek-r1 is no longer a callable model

R1 used to be the reasoning model you called by name. That door is closed. According to our live test, requesting deepseek-r1 returns an HTTP 400 stating the supported model names are deepseek-v4-pro or deepseek-v4-flash. R1's reasoning capability did not vanish; it was folded into V4. When we called the deepseek-reasoner alias, the response carried reasoning_tokens in its usage object while still reporting deepseek-v4-flash as the served model. So you still get reasoning behavior, you just stop addressing it as R1.

If your code still hard-codes model: "deepseek-r1", it is failing in production right now, not degrading quietly. Swap it for deepseek-v4-flash for the reasoning-capable cheap tier, or deepseek-v4-pro for the heavier tier. The stale $0.55/$2.19 R1 rate should come out of your spreadsheet entirely, because there is no model left to attach it to.

Change 3: the cache-hit price was cut 10x on 2026-04-26

According to DeepSeek API Docs, on 2026-04-26 DeepSeek reduced the cache-hit input price to one tenth of its launch-day cache rate. That single cut is what brought the current V4-Flash cache-hit rate to $0.0028 per million tokens, against a cache-miss rate of $0.14. The gap between those two numbers is 50x, and it now applies automatically with no flag to set. For a deeper walk-through of the mechanism, see our cache discount breakdown.

Read the two ratios carefully, because they are easy to conflate. The 10x is the size of the price cut on 2026-04-26: the cache rate dropped to 1/10 of what it had been. The 50x is the standing gap you pay against today, cache-hit $0.0028 versus cache-miss $0.14. Both are real, they describe different things, and most competitor posts blur them into one.

First-hand evidence: we measured the cache cut on a live call

Documentation says the discount exists. We called the endpoint twice to watch it fire. We sent an identical 557-token prompt prefix to deepseek-chat and recorded the raw usage object on each call. On the second request, the API reported prompt_cache_hit_tokens: 512 out of 557 prompt tokens, so 92% of the prefix was served from cache. The served model came back as deepseek-v4-flash both times, which confirmed the alias resolution inside the same run.

That 512-of-557 hit translates to roughly a 10x drop on the input side of that single call, because the cached portion bills at $0.0028 per million instead of $0.14. The number is not a docs claim; it is the cache_warm run in our evidence pack, captured 2026-05-13. The same artifact records the realistic-call run at 41 input and 75 output tokens, which is where the cluster's "output is most of the bill" observation comes from. We re-ran the cache test to be sure the hit was repeatable, and it was.

Change 4: the V4-Pro reversion that didn't happen

This is the freshest entry in the changelog, and it corrects our own earlier reporting. The V4-Pro 75%-off rate of $0.435 input and $0.87 output was announced to expire on 2026-05-31, with a nominal list price of $1.74/$3.48 scheduled for 2026-06-01. We told readers to budget for that step-up. It never arrived.

According to DeepSeek API Docs, as re-verified on 2026-06-25, V4-Pro is still priced at $0.435 input and $0.87 output, with a cache-hit input rate of $0.003625. There is no promo countdown on the page and no sign of the $1.74/$3.48 reversion. So treat $0.435/$0.87 as the standing live price, full stop. The $1.74/$3.48 figure is best understood as a nominal list price that was published but never billed; if you still have it in a spreadsheet as "what V4-Pro will cost after June," delete it.

When we called deepseek-v4-pro directly, the served model came back as deepseek-v4-pro, not an alias redirect, at about 2.64 seconds versus roughly 0.7 seconds for V4-Flash. For most workloads the cost difference matters more than the latency. At the standing rate, V4-Pro input costs about 3x V4-Flash ($0.435 versus $0.14), and that 3x is the multiple to plan around. There is no looming cliff to migrate ahead of.

For context, SiliconFlow mirrors the official V4-Flash rate at $0.14/$0.28 but charges $1.74/$3.48 on the Pro tier, so a third-party host can cost roughly 4x the official rate on V4-Pro. We measured the OpenRouter-routed comparators separately and disclose that routing; US-routed providers buy data residency, not a cheaper model.

FAQ

Why does deepseek-chat cost less than the old articles say? Because the name now resolves to deepseek-v4-flash. When we called deepseek-chat, the API returned served_model deepseek-v4-flash, which bills at $0.14 input and $0.28 output per million. The legacy V3 rate of $0.27/$1.10 that some 2025 posts quote is stale.

Can I still call deepseek-r1 on the official API? No. Requesting deepseek-r1 returns an HTTP 400 stating the supported names are deepseek-v4-pro or deepseek-v4-flash. R1 was folded into V4; the deepseek-reasoner alias gives you reasoning behavior while serving deepseek-v4-flash.

What exactly changed on 2026-04-26? DeepSeek cut the cache-hit input price to one tenth of its launch-day cache rate. That brought the V4-Flash cache-hit rate to $0.0028 per million, which is 50x below the $0.14 cache-miss rate. We verified the cache fires automatically on a repeated 557-token prefix.

Didn't the V4-Pro promo expire on 2026-05-31? It was announced to, but it didn't. Re-verified against official docs on 2026-06-25, V4-Pro is still $0.435 input and $0.87 output, with no promo countdown and no reversion to the nominal $1.74/$3.48 list price. Price V4-Pro at $0.435/$0.87, about 3x V4-Flash.

Is V4-Pro worth the premium over V4-Flash? At roughly 3x the input rate, choose V4-Pro when a task needs its heavier reasoning and the served model matters. For most workloads V4-Flash already carries reasoning behavior at $0.14/$0.28, so default to Flash and reserve Pro for the cases that justify the multiple.

This is part of the DeepSeek API pricing hub, which tracks every rate change in one place.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing. Last verified: 2026-06-25.

DeepSeek's API changed in four ways through mid-2026: deepseek-chat and deepseek-reasoner now resolve to deepseek-v4-flash, deepseek-r1 stopped being…

Written by Fan Chuanyu