Watch the 48-second breakdown: where DeepSeek V4.1 Flash wins, where it loses, and which jobs to give it.
Last week my timeline filled up with people saying a model that costs 60 cents per million output tokens is beating models that cost $25. That is a big claim, so I went through the launch notes, the Hacker News threads, X, and four YouTube reviews to see what is actually true.
Short version: DeepSeek V4.1 Flash is a very cheap open model that is excellent at long coding and terminal tasks. It is not better than Claude at everything. Here is the full picture.
What Is DeepSeek V4.1 Flash?
It is DeepSeek’s newest model, released on September 10, 2026. On October 17 they pushed a small update called V4.1-Flash-1017. That is the version you get on their website, app, and API today.
A quick naming warning:
- There is no V4.1 Pro yet. Only Flash exists. For the earlier generation, see our DeepSeek V4 features guide.
- The API name is
deepseek-flash. - Older names like
deepseek-v4-flashstill work, but they now send your request to V4.1 Flash.
What Changed Compared to the Old Version?
| Change | What it means for you |
|---|---|
| Much bigger model (552B parameters, up from 284B) | More brainpower, but harder to run at home |
| Reads images natively | You can send screenshots and it can look at its own app |
| 1 million token context | Fits a whole codebase |
| Tiny memory use for long text (about 890 bytes per token) | A full 1M-token chat needs under 1 GB of cache |
| MIT license | Free to use, change, and sell with |
| 1017 update | Shorter answers, faster finishes, better agent scores |
Why it works: the model reads your prompt using a small part of itself (about 8B parameters) and writes the answer with a slightly bigger part (about 16B). That keeps it fast and cheap even though the total model is huge.
DeepSeek V4.1 Flash Pricing
| Per 1M tokens | Off-peak | Peak (double) |
|---|---|---|
| Input already cached | $0.003 | $0.006 |
| Input (new) | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Peak hours are Monday to Friday, 01:00-04:00 and 06:00-10:00 UTC. Everything else is off-peak.
For comparison, the launch coverage lists Claude Opus 5 at $5 input and $25 output, and GPT-5.6 Sol at $4 input and $20 output. One developer on Hacker News said a whole project rarely costs more than $2-3 because cached input is so cheap. Another reviewer spent about $1.70 on six big 3D and web builds.
One catch: prices on other hosts like OpenRouter or Fireworks can differ, and DeepSeek’s own provider on OpenRouter may keep your prompts for training. Check the privacy line if your code is private.
DeepSeek V4.1 Flash Benchmarks vs Claude Opus 5 and GPT-5.6 Sol
Benchmark names mean nothing until you know what the test actually asks the AI to do. This is the official table for the 1017 version, with each test explained. Every score is DeepSeek’s own.
Where DeepSeek V4.1 Flash wins
| Test | What the AI has to do | This matters if you want to… | DeepSeek 1017 | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| DeepSWE | Fix real tasks in real open-source projects (TypeScript, Go, Python, JavaScript, Rust). Hidden tests check the finished code. | Fix bugs or add features in a project you already have | 77.0 | 74.0 | 73.0 |
| Automation-Bench | Finish multi-step business work across many apps (sales, support, finance, HR) without breaking company rules. Graded on the final result. | Automate boring office workflows | 60.5 | 50.3 | 45.8 |
| Agents’ Last Exam | Do real professional jobs on a real computer across 55 industries. Graded by code, not by opinion. | Hand over long, messy work tasks | 35.4 | 28.6 | 26.7 |
| Apex Bench | Solve very hard competition-level math problems. | Do heavy math and logic | 72.1 | 65.8 | 54.3 |
Where Claude Opus 5 is still ahead
| Test | What the AI has to do | This matters if you want to… | DeepSeek 1017 | Claude Opus 5 |
|---|---|---|---|---|
| Humanity’s Last Exam (no tools) | Answer extremely hard expert questions from memory, with no web search. | Ask tough knowledge questions without search | 42.6 | 56.3 |
| Terminal-Bench 4.0 | Work in a command line like a developer. Set up tools, fix system problems, run programs. 66 tasks, long time limits. | Do server setup, DevOps, and scripting jobs | 42.3 | 51.8 |
| ProgramBench | Rebuild a whole program when it only gets the finished app and its manual, with no code. | Build something from scratch with no guide | 29.1 | 37.0 |
| NL2Repo-Bench | Build a full, installable Python library from one written brief, starting with an empty folder. | Say “build me this app” and get a working project | 68.2 | 75.3 |
Basically a tie (all tested with tools)
| Test | What the AI has to do | DeepSeek 1017 | Claude Opus 5 |
|---|---|---|---|
| Chartography | Read and understand professional charts. | 83.0 | 84.0 |
| BabyVision | Basic seeing skills like spotting differences, mazes, and pattern puzzles. | 92.5 | 94.1 |
| ZeroBench | Extremely hard picture-based reasoning questions, with five tries allowed. | 51.0 | 52.0 |
Which model for which job
| Your job | Better pick | Why |
|---|---|---|
| Fixing bugs in an existing project | DeepSeek V4.1 Flash | Wins DeepSWE and costs far less |
| Business and office automations | DeepSeek V4.1 Flash | Wins Automation-Bench and Agents’ Last Exam |
| Building a whole app from a brief | Claude Opus 5 | Ahead on NL2Repo and ProgramBench |
| Server setup and command-line work | Claude Opus 5 | Ahead on Terminal-Bench 4.0 |
| Hard knowledge questions with no search | Claude Opus 5 | 56.3 vs 42.6 |
| Reading charts and screenshots | Either one | Scores are almost the same, so pick the cheaper one |
The 1017 update is a real jump over the September version. Terminal-Bench 4.0 went from 31.2 to 42.3, DeepSWE from 74.2 to 77.0, and Humanity’s Last Exam from 36.8 to 42.6.
Against GPT-5.6 Sol it wins most rows, and loses on Humanity’s Last Exam (42.6 vs 44.5) and ZeroBench (51.0 vs 53.0). Against Kimi K3 and GLM 5.3 it leads on almost everything in the table.
So the pattern is clear. It is strong at fixing real bugs and automating long jobs. It trails Claude on raw knowledge without tools and on the hardest terminal and repo-building tests.
One fine print line matters. DeepSeek ran Terminal-Bench 4.0 inside the Claude Code framework, DeepSWE inside Mini-SWE-V2, and ZeroBench with five tries. Scores move a lot when the agent tool changes, so test it in your own setup.
What People Are Saying
On X: most posts are about price. One developer ran a rocket-launch 3D scene and reported $0.78 for DeepSeek versus $6.10 for GPT-6.1 Sol. A Chinese developer said its motion-graphics work is about one notch below Claude. DeepSeek’s own announcement got 6.4M views, and OpenCode made the model free for a while.
On Hacker News: the top comment called it “a really strong model”. People use it as a cheap backup when Claude or Codex limits run out, or for night-time bug triage. Some doubt it beats Kimi K3 on general knowledge, and some argue 552B is not really “Flash” anymore.
In reviews:
- Speed: 175-290 tokens per second.
- Long builds can still take 18, 31, or even 80 minutes because the model thinks for a long time.
- One reviewer’s own test showed slightly worse coding and tool calling than the previous Flash, but better document search and instruction following.
- Output quality was mixed. Good dashboards and a working walk-through 3D house, but a weak boat game and a weak Blender car.
Best Use Cases
- Coding agents on a budget: let a pricier model plan, and let DeepSeek do the long implementation work. Several developers do exactly this. Another long-coding contender worth comparing is in our Meta Muse Wonder Full review.
- Repeating the same big context: cached input at $0.003 makes long projects cheap.
- Security and code review: one security team called it their best hacking model, and some users say it helps with tasks other models refuse.
- Building from screenshots: it can look at a screenshot, fix the page, and check again.
- Nightly or batch jobs: run them off-peak for half the price.
When to Skip It
- You need top scores on hard knowledge questions (42.6 vs Claude’s 56.3).
- You want the hardest terminal and repo-building tasks solved reliably.
- You cannot wait through long thinking times.
- You want to run it on a gaming PC.
Can You Run It at Home?
The weights are 510 GB. About 200 GB of that is lookup tables that can sit on an SSD or in normal RAM. People report these setups:
| Setup | Speed |
|---|---|
| Mac Studio M3 Ultra 256 GB (2-bit) | about 9.5 tokens/sec |
| Dual RTX 3090 + 128 GB RAM (2-bit) | about 30 tokens/sec |
| Two RTX Pro 6000 cards + about 300 GB RAM | 240-254 tokens/sec shared across several chats |
Regular llama.cpp cannot load it yet, so you need a special build. If you only have a normal gaming card, use the API instead.
Frequently Asked Questions
Is DeepSeek V4.1 Flash free?
The weights are free under the MIT license. The official API is pay-per-token, with the prices above.
Is DeepSeek V4.1 Flash better than Claude Opus 5?
On four agent tests (DeepSWE, Automation-Bench, Agents’ Last Exam, Apex Bench) it scores higher in DeepSeek’s own table. On Humanity’s Last Exam, Terminal-Bench 4.0, ProgramBench, and NL2Repo-Bench Claude is clearly ahead.
What is the difference between V4.1 Flash and V4 Pro?
V4.1 Flash beats the old V4 Pro on coding and agent tests and costs far less. V4 Pro still scored higher on GPQA and the text-only HLE set. DeepSeek has not released a V4.1 Pro yet.
What is the 1017 version?
A small update on October 17 that gives shorter answers, faster task completion, and stronger agent results.
Can it read images?
Yes. Image input is built in (JPEG, PNG, GIF, WebP).
My Take
If you write code with an AI agent, DeepSeek V4.1 Flash is worth testing this week because the price is so low. I would keep Claude for planning and hard reasoning, and send the long, repetitive implementation work to DeepSeek. Always check the first few results yourself, since scores change a lot depending on which agent tool you use.
