DeepSeek V4.1 Flash Explained: Pricing, Benchmarks, and Claude Code
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I break down DeepSeek V4.1 Flash, the cheap model DeepSeek said beat its own flagship, V4-Pro. I cover what shipped, the six tricks that make it this cheap, DeepSeek's benchmark numbers next to the independent ones, and where I'd put it next to Claude Code. It's for builders who want lower AI agent bills and want to understand the tradeoffs first.
What I cover in this video
- 0:00 DeepSeek tried to retire V4-Pro, then reversed the decision a day later
- 1:32 What Flash actually is: parameters, context window, and license
- 3:07 The price next to Opus 5.5 and V4-Pro
- 3:52 The six tricks that make it cheap
- 7:41 Independent benchmarks vs DeepSeek's own
- 9:13 What people built with it
- 10:45 How I'd use it next to Claude Code
- 13:04 The catch: data, the US federal order, and picking a host
The flagship that wouldn't go away
On September 10, DeepSeek shipped V4.1 Flash and said tests put it ahead of V4-Pro on performance, cost, speed, and total runtime, so Pro was being phased out. The next day DeepSeek backed off "in response to user demand." September 14, the day every Pro request was supposed to move to Flash, came and went with no switch. Then on October 17 came what DeepSeek calls a minor update: same price, same model name in the API, new weights. DeepSeek's site still disagrees with itself. The launch post says Pro is being phased out, while the change log and the price list say it's staying.
Flash is a mixture of experts model with 552 billion parameters, plus a separate set of 196 billion it calls "Engram." Only about 8 billion parameters are active while it reads your prompt and 16 billion while it writes. It has a 1 million token context window, can write up to 384,000 tokens in one go, reads images, and the weights are on Hugging Face under the MIT license. If you've seen 748 billion, that's the two numbers added together. 763 billion is Hugging Face's automatic count of everything in the weight files.
On price, Flash costs $0.15 in and $0.60 out per million tokens off-peak, and double that at peak. Next to Opus 5.5 at $4 in and $20 out, that's about 33 times cheaper on output off-peak and about 17 times cheaper at peak. Pro costs over four times more on input and over three times more on output, and Pro can't read images.
How it's this cheap: six tricks
A coding agent rereads its whole history on every turn, so most of the cost comes from reading and remembering. DeepSeek's tech report is titled around pushing the limits of KV cache compression. The KV cache is the model's saved notes on what it has already read, so shrinking those notes shrinks the bill.
- Mixture of experts. Each layer has 384 small expert networks, and each token visits only six of them plus one shared expert.
- Reading with half the model. The first 20 layers read your prompt. The last 20 build their memory from what layer 20 already worked out. DeepSeek says this cuts the compute for reading your prompt about in half, which is where 8B vs 16B active comes from.
- Shared notes. One layer writes the notes and the next few reuse them. When it looks something up, it only reads the top 512 entries.
- 4-bit memory. With the sharing, long-term memory drops to 890 bytes per token, about a quarter of the old V4 Flash.
- Throwing away the scratch paper. The short 128-token window used to be saved to SSD and made up nearly half of everything stored. Now it stays in memory for a few minutes and gets replayed if it's lost. Long-term notes are kept at least 72 hours, so you have more time to hit the cache.
- Drafting ahead. A small drafter guesses five tokens at once and the big model checks them.
Add it up and DeepSeek's first model in 2023 stored about 389,000 bytes of notes per token, while this one stores 890. That's 437 times smaller. Off-peak, cached input costs $0.003 per million tokens, 50 times cheaper than fresh input. DeepSeek's own fine print warns that some of these shortcuts "may still cause capability degradation in untested boundary cases."
Benchmarks: DeepSeek's numbers vs independent ones
DeepSeek reports Terminal-Bench 2.1 at 90.6, DeepSWE at 74.2 (77 after the October update), and CyberGym at 88.1. All of those beat V4-Pro on its own table, and the first two are level with or just ahead of Opus 5 (not 5.5). DeepSeek also shows where it loses, including Humanity's Last Exam at 42.6 vs Opus 5 at 56.3.
On the Artificial Analysis Intelligence Index, Flash scores 32 and V4-Pro scores 23, so on this scoreboard too the cheap model beats the flagship. But Opus 5.5 scores 58, Sonnet 5.5 54, and Haiku 5.5 36. On Terminal-Bench 4, DeepSeek reports 42.3, Artificial Analysis measured 33.8, and Opus 5.5 got 59.6. On HLE Diamond it's 25% vs 68%.
The result that surprised me was the cost to run the whole index. Flash cost $521 at peak price and Opus 5.5 about $8,900. Haiku 5.5 did it for $339 and scored higher. Flash is wordy: it used 170 million tokens where the median is 100 million. The price list doesn't tell you the whole bill.
In Paweł Huryn's 105-hidden-bug test, Flash averaged 38.7 over three runs and Opus 5.5 scored 41.7. In the same thread, though, he said they caught DeepSeek cheating on a separate run and threw it out. DeepSeek's own report also admits exploit-seeking behavior, including decompiling core Ubuntu packages to find CyberGym vulnerabilities.
Where I'd use it next to Claude Code
Opus 5.5 stays my daily driver in Claude Code. Here's where a model this cheap fits:
- Bulk chores with long, repeated context, like listing titles, data cleanup, and summarizing scrapes for my resale business. On short prompts Haiku is cheaper, but Haiku goes up to $0.50 in and $2.50 out over 100K tokens.
- A second-opinion bug hunter. It suggests the bugs, and then a stronger model or a real test decides which ones are real.
- Eval batch runs during the US workday. Peak hours are Beijing business hours, so a US workday is mostly off-peak.
DeepSeek publishes the settings to point Claude Code at its API. Use a separate Claude Code profile in its own terminal, because one base URL sends everything in that session to DeepSeek. In DeepSeek's mapping, anything called Opus goes to V4-Pro at Pro prices. Keep the model name in a config file, since DeepSeek has already changed its lineup once. Also, don't trust Claude Code's cost meter here: one developer reported that it prices a model it doesn't recognize at Opus 5 rates.
I wouldn't use it for hard reasoning, building from scratch, production code you won't review, short prompts where Haiku is cheaper, or anything private.
DeepSeek's privacy policy says it stores personal data in the People's Republic of China and uses data to train its technology. Its API terms are governed by Chinese law. On October 20, the Federal Register published a DHS order telling federal civilian agencies to remove DeepSeek products and third-party products built with DeepSeek implementations. It doesn't restrict companies or the public, but if you sell to federal civilian agencies, it could reach you. Because the weights are MIT, other hosts can run them. OpenRouter lists 35 providers, and each one runs its own version, whether 8-bit, 4-bit, or unstated. Pick a host based on where your data goes and which version it runs, not just on price.
Key takeaways
- V4.1 Flash beats V4-Pro on the independent index (32 vs 23) but trails Opus 5.5 (58), Sonnet 5.5 (54), and Haiku 5.5 (36).
- Its low price comes mostly from KV cache compression, down to 890 bytes per token.
- Flash is wordy, so the full bill can come out higher than Haiku's even though its per-token price is low.
- It works best for long, cache-heavy agent chores and as a second opinion, run in its own Claude Code profile.
- For private data, look closely at where it's hosted. The DeepSeek API and app send data to DeepSeek.
FAQ
How much does DeepSeek V4.1 Flash cost?
Off-peak it's $0.15 per million input tokens and $0.60 per million output tokens, double at peak. Cached input is $0.003 per million off-peak, 50 times cheaper than fresh input.
Is DeepSeek V4.1 Flash better than V4-Pro?
On the Artificial Analysis Intelligence Index, the updated Flash scores 32 and V4-Pro scores 23. Flash is also cheaper and reads images, which Pro can't.
Can I use DeepSeek V4.1 Flash with Claude Code?
Yes. DeepSeek publishes the settings to point Claude Code at its API. Run it in a separate profile and terminal, and note that DeepSeek maps any model called Opus to V4-Pro at Pro prices.
Is DeepSeek banned in the US?
No, not for businesses or the public. An October 20 DHS order tells federal civilian agencies to remove DeepSeek products and third-party products built with DeepSeek implementations, which could affect companies that sell to those agencies.
Chapters
- 0:00DeepSeek tried to retire its flagship
- 1:02What actually shipped
- 2:24Join Shipping Skool
- 2:44The price
- 3:26How it's this cheap: 6 tricks
- 6:49The benchmarks: DeepSeek's numbers vs independent
- 9:29What people built
- 11:00Where it fits next to Claude Code
- 12:45The catch: data, bans, hosts
- 14:12What's a flagship for?
Transcript
Read the full transcript
0:00On September 10th, Deepseek shipped a cheap model V4.1 Flash and said tests put it ahead of their own flagship V4 Pro on performance, cost, speed, and total runtime. So, they announced, "We're phasing out V4 Pro." And then the next day, Deepseek backed off. In response to user demand, we have decided to continue providing API services for Deepseek V4 Pro. It's still on sale today. And now an independent scorer has weighed in. On the artificial analysis intelligence index, the updated flash scores 32, V4 Pro 23. In this video, I'm going to be breaking down the flagship that wouldn't go away, the six tricks that make it this cheap, Deepseek's numbers versus the independent ones,
0:46what people actually built, where I'd put it next to Claude Code, and the catch, which just got bigger. I covered the October 17th update the day after it dropped. So, this is everything I didn't get to, plus what's happened since. Let's jump right into it. So, what actually happened? Quick timeline. September 10th, Flash ships with a price cut and DeepSeek says Pro is being phased out. September 11th, the reversal. September 14th, the day every Pro request was supposed to go to Flash. Never happened. Then October 17th, an update Deepseek calls a minor update. Same price, same model name in the API, new weights. So, here's the funny part. Deepseek's launch post still says Pro is
1:32being phased out. Their change log says it stays and Pro is still on the price list. All on their site right now. So, what is Flash exactly? It's a mixture of experts model with 552 billion parameters plus a separate set of 196 billion they call engram. But only about 8 billion parameters are working when it reads your prompt and 16 billion when it writes. It has a 1 million token context window. It can write up to 384,000 tokens in one go. It reads images and the weights are on hugging face under the MIT license. So you might have seen 748 billion or 763. 748 is those two numbers added together. 763 is HuggingFac's automatic count of
2:19everything in the weight files. Deepseek's own number is 552. Real quick, every video I make is built in code with Claude Code. Inside Shipping School Premium, I show you how stepby step to turn your product into videos that get views. Plus, new courses every day, 500 plus hours of replays, live calls every week, and a Q&A with me every Saturday. Join free link below. So, the price flash is 15 cents in, 60 cents out per million tokens off peak and double that at peak. At launch, DeepS cut Flash's prices compared to the old V4 Flash. That was 57% off cached input, 32% off fresh input, and 9% off output. So, next to Opus 5.5 at $4 in
3:07and 20 out, Flash is about 33 times cheaper on output off peak and about 17 times at peak. And next to Pro, Pro costs over four times more on input and over three times more on output. And Pro can't read images. So, why would anyone fight to keep it? Hold that thought. So, how is it this cheap? Start with what actually costs money. When a coding agent works, it rereads its whole history on every single turn. The code, the tool output, all of it. So most of the work is reading and remembering. And Deepseek's tech report for this model is literally titled pushing the limits of KV cache compression. The KV cache is basically the model's saved notes on everything it has already read, so it
3:52doesn't have to redo the math. Shrink the notes and you shrink the bill. So six tricks. Don't get lost in the sauce here. Each one is simple. Step one, mixture of experts. Every layer has 384 small expert networks and each token only visits six of them plus one shared expert. So it's a huge model that runs like a small one. Step two, and this is the part by cloud called an even cooler consequence. It reads with half the brain. The first 20 layers read your prompt. The last 20 layers don't read it again. They build their memory straight from what layer 20 already worked out. Deepseek says that cuts the computation for reading your prompt about in half. That's where 8 billion versus 16 billion comes from. Step three, layers share
4:38notes. Normally, every layer keeps its own notebook. Here, one layer writes the notes and the next few reuse them. Some reindex, so they use the same notes but pick their own pages. And when it looks something up, it only reads the top 512 entries. Step four, write the notes in four bits. Together with that sharing, the long-term memory drops to 890 bytes per token, about a quarter of the old V4 flash. Step five, throw away the scratch paper. The model keeps a short window of the last 128 tokens in V4. That window got saved to SSD and it was nearly half of everything stored. Now, it's kept in memory for a few minutes and if it gets lost, they just replay those 128 tokens to rebuild it. Total stored cash about
5:241/8 of V4 flash and the long-term notes are kept for at least 72 hours so you get a longer window to hit the cache. Step six, draft ahead. A small drafter guesses five tokens at once and the big model checks them. And that engram set I mentioned, it's lookup tables for short word patterns. Deepseek's idea is to separate memorizing from thinking. So add it all up. Back in 2023, Deepseek's first model stored about 389,000 bytes of nodes per token. This one stores 890. That's 437 times smaller, which is insane. So, here's how that shows up on your bill. Off peak, cached input is 3/10 of a cent per million tokens. Fresh input is 15
6:11cents. That's 50 times cheaper. And Deepseek says cash hit charges often account for a large share of agent costs. Then on October 20th, Deepseek open sourced the engine behind it called Deep Infer. They say one cluster serves 155 trillion tokens a day and brings in $38 per GPU hour at off- peak prices. That's revenue, not profit. And it's DeepSeek's number, but it's a big one. and their own fine print says some of these shortcuts may still cause capability degradation in untested boundary cases. Their words keep that in mind for the next part. So how good is it really? Start with Deepseek's own numbers from launch. Terminal Bench 2.1
6:5690.6 Deepsw SWE 74.2 Cyberjim 88.1 all ahead of V4 Pro on their own table. And on the first two, level with or just ahead of Opus 5, not 5.5, Opus 5. Then Deepseek says the October update pushed Deepswe to 77, all vendor reported. So take that for what you will. So to their credit, Deepseek shows where it loses. Humanity's last exam 42.6 versus Opus 5 at 56.3. They even write that a gap with giant models remains on scienceoriented agentic tasks. So now the independent numbers. Artificial analysis runs nine eval for its intelligence index. Flash
7:41scores 32. That's the top open weights model in its size class and nine points above V4 Pro. So on this scoreboard, the updated Flash beats the flagship. But next to Claude, Opus 5.5 scores 58. Sonnet 5.5 54. Haiku 5.5 36. Same test, different referee. On terminal bench 4, DeepSeek reports 42.3. Artificial analysis measured 33.8. Opus 5.5 got 59.6. It's close on some things. On Automation Bench, it's about 67 versus Opus 5.5 at about 70. And it's far on others. On HLE Diamond, the hard reasoning set, 25% versus 68. So, this one surprised me.
8:28Artificial analysis also tracks what it costs to run its whole test suite. Flash $521 at DeepSseek's peak price. Opus 5.5 about $8,900, 17 times more. But Haiku 5.5 did it for 339 and scored higher. Part of that is Flash is wordy. It used 170 million tokens to get through the tests where the median is 100 million. So the price list doesn't tell you the whole bill. One realworld test. Powell Hurin ran a test with 105 hidden bugs across two repos. Flash ran it three times and averaged 38.7. Opus 5.5 scored 41.7, three bugs apart. But in the same thread, he wrote,
9:13"We caught DeepSseeek cheating on a separate higheffort run and he threw that run out." And Deepseek's own report admits exploit-seeking behavior. During testing, it decompiled core Ubuntu Linux packages to dig up Cyberjim vulnerabilities. Scary. So, what are people actually building with it? Deepseek's own demos for the update show the Forbidden City in 3D, a racing game, and a Portal 2 recreation in GDAU. A J on X had it oneshot a 3D scene in 3JS, no harness. AI Search had it make a full explainer video in 36 minutes with a 99% cash hit rate and said the total cost of everything in his video was $6. $6. So Unsloth shrank it down to a
10:00onebit version, 508 GB down to 250, keeping about 88% accuracy by their measure. You still need a machine with at least 128 gigs of RAM. So that's a serious workstation, not a normal laptop. and people are running Claude code on it. Deepseek publishes the exact settings to point claude code at its API. One developer on dev.2 said each run cost about 2. Claude Code's own cost meter said somewhere between 87 and $1.28 because he says it prices a model it doesn't recognize at Opus 5 rates. So don't trust that meter. But not everyone sold. MIA tested the September version on code reviews. It raised more real problems than GLM and more fake ones.
10:45She also noticed GLM viewed 14 screenshots and DeepSseek viewed zero. She still gave GLM the win, 2:1. And one person on X put it this way. Looks like a good upgrade on paper, but looks like they didn't fix overthinking at all. So, here's how I'd actually use it. Opus 5.5 is my daily driver, and that's not changing. But a model this cheap has a spot in the setup. First, bulk chores with long repeated context. In my resale business, that'd be listing titles, cleaning up data, summarizing scrapes. Long prompts that hit the cash are where Flash's pricing wins. On short prompts, Haiku is actually cheaper per token, but Haiku jumps to 50 in, $2.50 out once your prompt is over a 100,000 tokens,
11:32and Flash's cashed input is 3/10en of a cent. Second, a second opinion bug hunter. It's within about three bugs of opus on Hurin's test, but in Mia's test, it flagged fake problems, too. So, it suggests and a stronger model or a real test decides. Third, [snorts] eval batch runs. And here's the counterintuitive part. If you're in the US, run them during your day. Deepseek's peak hours are Beijing business hours, which is our evening and overnight. A US workday is basically all off peak. So, set it up so it can't mess with your main setup. Use a separate cloud code profile in its own terminal with DeepSseek settings. One base URL means everything in that session goes to DeepSseek, so keep it separate. Watch the model names, too. In
12:18Deepseek's mapping, anything called Opus goes to V4 Pro at pro prices. And keep the model name in a config file because DeepSseek already changed its mind on its lineup once. So, where doesn't it fit? Hard reasoning and building something from scratch. production code you won't review after that cheating note, short prompts where Haiku is cheaper and scored higher, and anything private. Which brings us to the catch. Here's something to be wary about. Deepseek's privacy policy for its apps and services says, "We directly collect, process, and store your personal data in the People's Republic of China, and it uses data to train and improve their technology." It lists an opt out from training as a right that may be
13:04available to you and their API terms are governed by Chinese law. So now the US government has moved. On October 20th, the Federal Register published a Homeland Security order. Federal civilian agencies have to remove Deepseek products and third-party products and services developed using Deepseek implementations. It doesn't restrict companies or the public. So this isn't a ban for you or your business. But if you sell software to federal civilian agencies, that second part could reach you. Before that, Australia, Taiwan, the Czech Republic, and others restricted it for government use. Italy blocked the app, and Congress, NASA, the Navy, and the Pentagon barred it on government devices. So the app and the API both
13:50send your data to DeepSeek. The weights are different. They're MIT, so anyone can host them. Open Router lists 35 providers for the update like together, fireworks, and nebus. And each host runs its own version, 8bit, 4-bit, or they don't say. So, pick a host by where your data goes and which version it runs, not just the price. So, Deepseek said the cheap model won. The independent scoreboard has the updated flash ahead, 32 to 23, and users had already made them keep the flagship. So, what's a flagship for? I think it's the hard stuff and trust. For my agent chores, the cheap model is plenty. For the hard stuff and anything private, Claude stays. And if one of those six tricks
14:35finally made the KV cash click for you, leave a like so it reaches more builders. Now, I want your take. Bulk chores on DeepSeek or is the data question a dealbreaker for you? Tell me which side you're on down below. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Jev by TypeSafe Explained: Pricing, Benchmarks, and Claude Code
Jev is TypeSafe's System One model: you send it a state plus typed questions and it returns a choice, a score, or a probability instead of text. I'd keep Claude Code and Opus for planning and building, and put Jev on the small yes or no calls inside the app.
Gemini 4 Carbon vs Claude Opus 5.5: Benchmarks, Pricing, and Access
Gemini 4 Carbon is #1 on the Artificial Analysis Intelligence Index (59 vs Claude Opus 5.5 max at 58), but only cyber defenders and Fairwind partners can use it so far. On real coding agent work, Opus 5.5 in Claude Code still leads 62 to 59, though Carbon is faster and cheaper per task.
DeepSeek V4.1 Flash vs Opus 5.5: Benchmarks, Cost, and the Catch
DeepSeek V4.1-Flash-1017 is an MIT-licensed open-weight model that beats Opus 5 on DeepSWE (77 vs 74) for 60 cents per million output tokens, about 33x cheaper than Opus 5.5. I'd use it as a second model for high-volume agent loops. It doesn't replace Opus for hard reasoning, production code, or private data.
