DeepSeek Flash Pricing: V4.1 Cut 31%, Measured

DeepSeek Flash pricing changed on 10 September 2026, and most WordPress chatbots running on DeepSeek got the new rate without anyone touching a setting. That day DeepSeek released V4.1 Flash, retired V4 Flash, and cut the Flash rate card: cache-miss input from $0.44 to $0.30 per million tokens at peak, output from $1.32 to $1.20, cache hits from $0.014 to $0.006. The old model name, deepseek-v4-flash, was not switched off. DeepSeek routes it to the new model and bills it at the new price.

That is the kind of change that reads well in a changelog and leaves the practical questions open. Is the old name really the new model? Does V4.1 count tokens differently, the way two other new models did this month? What does a reply cost now, and does it change the verdict of our August analysis of DeepSeek’s peak-hour pricing, which found DeepSeek’s cheapest model had become more expensive than GPT-5.6 Luna for a retrieval chatbot? We measured it: 24 real calls to the DeepSeek API from a production WordPress server on Sunday 27 September 2026, with the same system prompt, retrieved documents and visitor questions we used for our OpenAI tests the day before.

• A retrieval chatbot reply on V4.1 Flash costs $0.77 per 1,000 off-peak and $1.53 at peak, with thinking disabled. On the same token counts, retired V4 Flash would have cost $1.26 and $2.52. That is a 31% cut on the bill, driven almost entirely by the input rate.
• The legacy name is the new model. Every deepseek-v4-flash call came back labelled deepseek-flash with the same system fingerprint, and it read from the cache that deepseek-flash calls had just written.
• V4 Pro costs 4.9 times as much for the same replies and, on our test, reused no cached input at all until a question repeated word for word.

The DeepSeek Flash rate card, before and after

This is DeepSeek’s Models & Pricing page as it read on 27 September 2026, in US dollars per million tokens, next to the V4 Flash rates that applied from 16 August until V4.1 shipped. Off-peak is exactly half of peak on every line.

Three details on that page matter more than the headline cut.

The model name. DeepSeek now asks you to call deepseek-flash. The page says the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, that the models behind them have been retired, and that their requests are served by V4.1 Flash and billed at the Flash price. The changelog uses the word “temporarily” for that routing. No end date is given.

V4 Pro did not go away. DeepSeek had signalled that V4 Pro would be phased out after 14 September. The 10 September changelog reverses that: “in response to user demand” V4 Pro service continues after 14 September with billing unchanged. Its rates are the same as in August.

Peak hours now exclude Chinese public holidays. Peak is still 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, but the footnote now adds “excluding Chinese public holidays”, which are off-peak in full. We come back to why that matters next week.

How we measured it

Rate cards are priced per million tokens; a chatbot is billed per reply. To turn one into the other you need to know what one reply actually sends, so we sent real ones. The test ran from this site’s WordPress server with the DeepSeek API key already stored in the MxChat settings, so the key never left the server.

• Prompt shape: the live MxChat system prompt (3,478 characters) plus six knowledge-base documents retrieved for each question, capped at 3,200 characters each, plus the visitor question. This is the retrieval (RAG) shape a WordPress site chatbot uses: thousands of tokens in, a few sentences out.
• Questions: A (WooCommerce order lookup), B (whether Pinecone is required), C (handing a visitor to a human on Slack or Telegram), then A again word for word to test the cache.
• Models: deepseek-flash, the legacy deepseek-v4-flash, and deepseek-v4-pro.
• Modes: thinking disabled ("thinking": {"type": "disabled"}, which is what MxChat sends to DeepSeek) and the API default, which is thinking on.

That is 3 models × 2 modes × 4 calls = 24 calls. All returned HTTP 200. Sunday is off-peak all day, and the whole test cost about three cents at off-peak rates; the account balance, which DeepSeek reports to the cent, did not move.

What 24 calls returned

The old name is the new model

The model field in every deepseek-v4-flash response read deepseek-flash, and the system_fingerprint was identical to the deepseek-flash calls. The stronger evidence is the cache. We called deepseek-v4-flash seconds after sending the same prompts to deepseek-flash, and its very first call read 5,888 cached tokens. DeepSeek’s cache is kept per model, so the simplest reading is that the two names share one model and one cache. If your site is set to deepseek-v4-flash, you are already on V4.1 Flash and already paying its price.

The flip side is the word “temporarily”. A name that is being routed can stop being routed. Changing the model to deepseek-flash costs nothing and removes the one way this change could break a working chatbot later.

No tokenizer surprise this time

Twice this month a new model shipped at an unchanged price with a tokenizer that counts more: Anthropic’s, which we covered in how Claude’s tokenizer changes your bill, and Grok 4.7. V4.1 Flash belongs to a new architecture family, so it was worth checking. It counted 6,088 tokens for prompt A; V4 Pro, from the previous family, counted 6,087. B and C differed by one token as well. For a chatbot the cut on the rate card is the cut on the bill.

One caveat we cannot remove: the retired V4 Flash can no longer be called, so the “before” price in this article applies V4 Flash’s rates to V4.1’s token counts. Given that V4 Pro and V4.1 Flash agree to within one token, we think that is a fair assumption, but it is an assumption.

What a reply costs on DeepSeek V4.1 Flash

The three new questions averaged 5,581 input tokens and 47 output tokens on V4.1 Flash with thinking off. That ratio, about 120 tokens in for every token out, is why the input rate decides the bill. Priced per 1,000 replies, with the cache hits the test actually got:

The GPT rows come from our measurement of the same three questions on 26 September, written up in what OpenAI’s cache writes do to a chatbot bill. Each vendor’s own token counts are used for its own rows.

Read against our August verdict, three things have moved.

• DeepSeek’s cheapest model is cheaper than GPT-5.6 Luna again, off-peak. In August V4 Flash off-peak cost 3.7% more than GPT-5.6 Luna for this workload. V4.1 Flash off-peak is about half the price of GPT-5.6 Luna with its default caching. At peak the two are close: $1.53 against $1.44 with caching on, or $1.17 with it off.
• GPT-6 Luna moved further. OpenAI’s 22 September release roughly halved Luna’s rates, to $0.10 input and $0.50 output. V4.1 Flash off-peak charges more on input ($0.15) and more on output ($0.60), so for a retrieval chatbot with caching turned off, GPT-6 Luna is the cheaper of the two by about a fifth. With OpenAI’s default caching on, which writes every new retrieval prompt at 1.25x, the two are within 4% of each other.
• Where DeepSeek wins is the cache-hit meter. V4.1 Flash bills cache hits at $0.003 per million off-peak, a third of GPT-6 Luna’s $0.01 cache read, and DeepSeek charges nothing to write the cache. Our exact-repeat call cost the equivalent of $0.075 per 1,000 replies. A bot with a large fixed prompt and short, repetitive questions will find DeepSeek cheaper than this table suggests.

How DeepSeek’s cache behaved on a retrieval bot

DeepSeek caches automatically. There is nothing to switch on, no write charge, and hits appear in the response as prompt_cache_hit_tokens. The question for a chatbot is how much of each new prompt is a repeat of something recently sent.

On V4.1 Flash, question B read 1,408 cached tokens and question C read 640. Both are the opening the three prompts have in common: the system prompt, and on B the start of the retrieved text as well. The exact repeat of A read 5,888 of 6,088. Every hit count we saw was a multiple of 64, so the cache matches in blocks and the last partial block of a shared opening is always billed at the miss rate.

V4 Pro behaved differently. Questions B and C read nothing from the cache, even though they began with the same 3,478-character system prompt as A, and only the word-for-word repeat hit (6,016 tokens). One run on one morning is not proof that V4 Pro never reuses a shared prefix, but it did not on this test, in either mode.

In money terms, the partial hits saved about 12% on V4.1 Flash ($0.87 to $0.77 per 1,000 off-peak). That saving grows with the size of your system prompt, and it is the reason to keep the system prompt fixed and first: anything that changes near the top of the prompt, such as a timestamp or the visitor’s name, pushes the shared opening down to zero. The same advice applies to OpenAI, and we measured the cost of getting it wrong in the caching article linked above.

Thinking is on by default, and it bills as output

Both DeepSeek models think before answering unless told not to. On a knowledge-base question that is mostly wasted: the answer is in the retrieved text, and the visitor wants it quickly.

The cost penalty is small because output is a small share of a retrieval bill; the latency penalty is not. V4 Pro with thinking on took 2.5 times as long to answer, for a shorter reply. MxChat already sends thinking: disabled to DeepSeek on chat, translation and content-generation calls. If you call the API from your own code, set it yourself, because leaving it out means thinking on.

Peak hours, and the week they disappear

DeepSeek’s peak windows are published in UTC and fall mostly outside a US audience’s working day. Our peak-hours breakdown has the full timezone table; the short version is below.

The new holiday exclusion is worth a calendar note this week. China’s National Day holiday for 2026 runs from Thursday 1 October to Wednesday 7 October under the State Council’s published schedule. Five of those days are weekdays that would normally carry 35 peak hours between them. Under DeepSeek’s footnote, all of them are off-peak. If you have a large reindexing, translation or summarisation job to run through DeepSeek, that week is the cheapest time this quarter to run it, at no cost to live chat traffic.

Two limits on that. DeepSeek’s page does not say how it treats the make-up working days China schedules around the holiday (for 2026, Sunday 20 September and Saturday 10 October), so do not count on a Saturday being off-peak or peak until DeepSeek says. And a holiday footnote can change; check the pricing page on the day.

What to do if your WordPress chatbot runs on DeepSeek

1. Change the model name to deepseek-flash. You already have the new model and price on the old name, but the old name is routed “temporarily”. If your plugin does not list deepseek-flash yet, the legacy name keeps working for now; recheck after each plugin update.
2. Keep thinking off for chat. It cost 7% more on Flash and made Pro 2.5 times slower on our test without a better answer.
3. Do not use V4 Pro for knowledge-base answers on price grounds. At 4.9 times the cost of Flash, and with no prefix reuse on our test, it is a model for agent work and hard reasoning. On the same three questions, Flash answered in 1.03 seconds on average against Pro’s 1.56 with thinking off.
4. Put the fixed part of the prompt first and keep it fixed. DeepSeek’s cache is free and automatic, but it only matches from the start of the prompt, in 64-token blocks.
5. Compare against your own ratio, not a headline. A retrieval bot is about 99% input. At that shape, GPT-6 Luna with caching off is about a fifth cheaper than V4.1 Flash off-peak; a bot with a big static prompt and repetitive questions tips the other way.
6. Put a spending limit on the key. Cheaper tokens do not change the risk of someone scripting your chat endpoint; see denial of wallet on WordPress chatbots for what that looks like.

MxChat supports DeepSeek alongside OpenAI, Anthropic, Gemini, xAI and OpenRouter, and the model is a setting rather than a rebuild. The MxChat documentation covers adding a DeepSeek key and choosing a model, and MxChat Pro adds the retrieval controls that decide how many input tokens each reply sends, which on every model in this article is the number that sets the bill.

Limits of this test

• 24 calls, three questions, one Sunday morning. Token counts are deterministic for a given prompt; cache behaviour and latency may vary with load.
• The retired V4 Flash could not be called, so its costs are its published rates applied to V4.1 Flash’s token counts.
• Answer quality was not scored. We note only that answer lengths were similar across models.
• All prices are DeepSeek’s own API. Third-party hosts of V4.1 Flash set their own rates.

FAQ

How much does DeepSeek V4.1 Flash cost?

Per million tokens: $0.15 input on a cache miss, $0.003 on a cache hit and $0.60 output off-peak; exactly double at peak ($0.30, $0.006 and $1.20). Peak is 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays.

Is deepseek-v4-flash still available?

The name is still accepted, but the V4 Flash model was retired on 10 September 2026. Requests to deepseek-v4-flash are served by V4.1 Flash and billed at the V4.1 price. In our test the responses came back labelled deepseek-flash and shared its cache. DeepSeek describes the routing as temporary.

How much cheaper is V4.1 Flash than V4 Flash?

Cache-miss input is 32% cheaper, output 9% and cache hits 57%. On a retrieval chatbot, where input is about 99% of tokens, the bill per reply fell 31%: from $1.26 to $0.87 per 1,000 replies off-peak before any cache hits, or $0.77 with the hits our test received.

Is DeepSeek cheaper than GPT-6 Luna for a chatbot?

For a retrieval chatbot, not quite. GPT-6 Luna charges $0.10 input and $0.50 output against V4.1 Flash’s $0.15 and $0.60 off-peak, so Luna with caching off was about 20% cheaper on our test. With OpenAI’s default caching on they were within 4%. DeepSeek is cheaper on cached input, so bots that resend the same long prompt do better on DeepSeek.

Does DeepSeek charge for thinking tokens?

Yes. Reasoning tokens are billed at the output rate, and thinking is on by default on both V4.1 Flash and V4 Pro. Send "thinking": {"type": "disabled"} to turn it off.