DeepSeek V4.1 Flash 1017: Benchmarks, Price, and a Real Coding Agent Test

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026. The number is the date: it is a checkpoint update to the V4.1 Flash model from September. DeepSeek’s model card calls it a “minor update” with “substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks,” plus shorter outputs and faster task completion. The weights are open under the MIT License.

On DeepSeek’s API it costs $0.30 per million input tokens and $1.20 per million output at peak hours, and half that off-peak. It replaced the old V4 Flash: requests to deepseek-v4-flash now go to 1017.

We ran it the way most people will use it, as a coding agent. Same task, same agent, three runs each for 1017, the September V4.1 Flash, MiniMax M3.1 Flash, GLM 5.3 Flash, Claude Haiku 5.5, and Claude Sonnet 5.5.

Model IDs. DeepSeek API: deepseek-flash (the old deepseek-v4-flash also works). OpenRouter: deepseek/deepseek-v4.1-flash-1017. Weights: deepseek-ai/DeepSeek-V4.1-Flash-1017. In Atomic Agent, pick the DeepSeek preset or OpenRouter, then the model with /model.

What is DeepSeek V4.1 Flash 1017?

DeepSeek V4.1 Flash 1017 is the October checkpoint of DeepSeek’s V4.1 Flash, a 552B-parameter mixture-of-experts model with a 1 million token context, up to 384K output tokens, image input, and MIT-licensed weights. It keeps the September architecture and adds more reinforcement learning, which DeepSeek says improves end-to-end project work.

From DeepSeek’s September launch post, pricing page, and the Hugging Face model card.

The parameter count needs a note. DeepSeek describes the model as a “552B-parameter MoE.” Hugging Face’s file metadata shows 763B parameters, most likely because the stored checkpoint also counts its large Engram memory tables. Both numbers describe the same download.

The naming is easy to mix up:

• V4.1 Flash 1017 is this release, October 17.
• V4.1 Flash without a date is the first checkpoint, September 10. Some providers still host it under that ID.
• V4 Flash is the older generation. DeepSeek retired it, and its API name now points to 1017.
• V4.1 Pro hasn’t been released.

DeepSeek also recommends a temperature and top_p of 1.0 and max_tokens of at least 256K, and ships no Jinja chat template. The model card points to a Python reference encoder and a Rust library, deepseek-recipe, instead.

Is DeepSeek V4 Flash the same model now?

On DeepSeek’s own API, yes. DeepSeek’s pricing page says the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp “are still accepted, but the corresponding models have been retired, their requests are served by the DeepSeek-V4.1-Flash-1017 model and billed at the Flash price.” If your code still calls V4 Flash, it already runs 1017.

Third-party hosts are a different story. On OpenRouter, deepseek/deepseek-v4-flash and deepseek/deepseek-v4.1-flash are separate listings served by other providers, so pick the dated ID deepseek/deepseek-v4.1-flash-1017 if you want this checkpoint.

V4 Flash or V4 Pro? DeepSeek’s API still lists deepseek-v4-pro (V4-Pro-0813) at $1.32 input and $3.96 output per million tokens at peak, about 4.4 times the Flash input price and 3.3 times the output price. DeepSeek currently lists it at its own price. DeepSeek said in September that tests by “multiple parties” put V4.1 Flash ahead of V4 Pro on performance, cost, and speed.

How good is DeepSeek V4.1 Flash as a coding agent?

It did well on our task. All three 1017 runs passed all 32 hidden tests, in 26 to 28 seconds each, with about 14 model calls per run. Only the September checkpoint was faster, and at $0.016 per task it costs close to Claude Haiku 5.5 ($0.011) and the September checkpoint ($0.014).

How we tested

We used the same task as in our Claude Sonnet 5.5 test: a small Python package with five planted bugs and two missing features (a settlement function and a command-line tool), plus 32 hidden tests the agent never sees. The prompt is one line: read the task file, complete it, run the tests until they pass. Each model ran three times in Atomic Agent v0.6.9 through OpenRouter on October 19, 2026, with approvals off and each model’s default settings, except the MiniMax low-effort runs, where we set reasoning effort to low.

We pinned each model to one provider so OpenRouter couldn’t route it to whichever host was cheapest at the moment. Both DeepSeek checkpoints ran on Together, which charges $0.30 input, $1.20 output, and $0.006 cached input, the same as DeepSeek’s peak price. DeepSeek’s own endpoint on OpenRouter was filtered out by our account’s data settings, so the off-peak price below is an estimate.

Results

Three runs per setup, averages shown. Token counts are the provider’s native counts from OpenRouter’s generation records. MiniMax M3.1 Flash ran on its free OpenRouter route, so its cost is estimated at MiniMax’s list price; all other costs are what OpenRouter billed.

Every setup solved the task in every run, so this table doesn’t rank the models by intelligence. It shows how much work each one does to get there, and that work is what you pay for in time and money.

DeepSeek took more steps than the Claude models (14 against 6 to 8) but each step was cheap and quick. Most of its input came from cache: about 93% of the roughly 316,000 input tokens per run were cache hits, which DeepSeek prices at $0.006 per million at peak. In an agent loop, where every step resends the conversation, that cache price matters more than the headline input price.

At DeepSeek’s own off-peak rates ($0.15 input, $0.60 output, $0.003 cached), the same token counts would cost about $0.008 per task. That’s our estimate from the token counts, not a measured bill.

Limits of this test. One small, well-specified Python task, three runs per setup, on one day. It shows cost and speed on a task every model can solve. It says nothing about hard, long, open-ended work, which is where DeepSeek’s benchmark gains are reported. We used Together, not DeepSeek’s own endpoint.

Download Atomic Agent to repeat this on your own repo: switch the model per run and keep the same tools and prompt.

How do I use DeepSeek V4.1 Flash in Claude Code or Atomic Agent?

Any agent that accepts an OpenAI-style or Anthropic-style API can use it. DeepSeek’s API speaks both formats, with https://api.deepseek.com for OpenAI-style clients and https://api.deepseek.com/anthropic for Anthropic-style ones such as Claude Code. Use deepseek-flash as the model name.

In Atomic Agent you have two routes: DeepSeek’s own API, which gets you the off-peak discount, or OpenRouter, where one key reaches DeepSeek and every other model in the table above.

Install and start the agent:

curl -fsSL https://atomicagent.io/install | sh
atomic-agent

On Windows, run irm https://atomicagent.io/install.ps1 | iex in PowerShell, or download the desktop app.

• DeepSeek API. In the TUI’s provider wizard, pick the DeepSeek preset. It fills in https://api.deepseek.com and saves your key as DEEPSEEK_API_KEY. The model list comes from DeepSeek’s server, so deepseek-flash shows up as soon as DeepSeek lists it. Details are in our models guide.
• OpenRouter. Choose Cloud models, then OpenRouter, paste your key, and pick deepseek/deepseek-v4.1-flash-1017. The full steps are in our OpenRouter guide.

Switch models later with /model. Because the tools and prompt stay the same, you can run the same task on DeepSeek and on another model and compare, which is how we built the table above.

If you’d rather keep most of the work on your own machine, Atomic Agent’s Fusion mode can use a cloud model like DeepSeek as the orchestrator, which plans and reviews, while local workers do the edits. Only the orchestrator calls the API, though it still reads your project files. Our docs note this direction suits tasks that split into many small parts.

Is 1017 better than the September V4.1 Flash?

On DeepSeek’s benchmarks, clearly: Terminal-Bench 4.0 rises from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and DeepSWE v1.1 from 74.2 to 77.0. On our small task, which both checkpoints solve every time, we saw no real difference in time or cost.

DeepSeek says 1017 brings “faster task completion, and shorter outputs than the previous version.” Our runs don’t show that on easy work. Average reasoning was lower for 1017 (858 against 1,563 tokens), but only because one September run reasoned for 3,351 tokens; the other two used about 670. Total output and steps were slightly higher for 1017. With three runs each, we read this as a tie.

The case for upgrading rests on harder work. If you call deepseek-flash on DeepSeek’s API, there’s nothing to decide: you already get 1017.

How much does DeepSeek V4.1 Flash cost per task?

In our test, $0.016 per task at the peak-equivalent price and an estimated $0.008 off-peak. DeepSeek’s pricing is $0.30 input, $1.20 output, and $0.006 for cached input per million tokens at peak, and exactly half off-peak, which covers most of the day.

From DeepSeek’s pricing page, October 19, 2026.

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else, including weekends, is billed off-peak. If you run batch agent jobs, schedule them outside those windows.

On OpenRouter, 1017 is served by many providers at different prices. About ten providers, including DeepSeek, Together, and Fireworks, charge $0.30 and $1.20; others charge less, and many of those are labelled FP8 or FP4. If you want a specific host, pin the provider in your request.

For scale: at our measured $0.016 per task, 1,000 similar agent tasks a month cost about $16, against about $183 for Claude Sonnet 5.5 and about $11 for Claude Haiku 5.5.

Is DeepSeek V4.1 Flash free?

Partly, in two ways. OpenCode offers a free variant for a limited time, and for DeepSeek API users, its new Search and Fetch tools are free through some agents until October 31. The API itself is paid.

• OpenCode’s free model. opencode/deepseek-v4.1-flash-free needs no API key, according to TechPilot’s test. It has a smaller window, 200K context and 32K output against 1M and 384K on the paid API, only works inside the OpenCode app, and has no published end date.
• Search and Fetch. DeepSeek opened its Search and Fetch APIs to all developers on October 17 at $0.75 per 1,000 calls each. For DeepSeek API users, calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex are free until October 31, 2026, 16:00 UTC.

How does DeepSeek V4.1 Flash 1017 score on benchmarks?

DeepSeek reports 77.0 on DeepSWE v1.1, 42.3 on Terminal-Bench 4.0, 60.5 on Automation-Bench, and 35.4 on Agents’ Last Exam, ahead of Kimi K3, GLM-5.3, and GPT-5.6 Sol on every agent test in its own table. Claude Opus 5 leads on Terminal-Bench, ProgramBench, NL2Repo, and HLE.

Vendor-published, from the Hugging Face model card. Not our tests. The GLM-5.3 HLE figure is on the text-only subset.

DeepSeek names its harnesses for two of these: Claude Code for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE. It also published a chart of DeepSWE score against cost per task, which The Decoder reads as about $0.20 per task for the 77.0 result. The rival models in DeepSeek’s table are not the newest from each lab, so treat the comparison as a snapshot.

For MiniMax’s newest Flash model, tested on the same agent task, see our MiniMax M3.1 Flash test, and for the two side by side, including which is cheaper at peak and off-peak hours, see DeepSeek V4.1 Flash vs MiniMax M3.1 Flash.

Can you run DeepSeek V4.1 Flash locally?

Only on a big machine. With 552B parameters, even aggressive community builds are hundreds of gigabytes. One of them, a 2-bit build for the ds4 engine, is a 341 GiB file designed to stream from SSD on a 128 GB Mac, and its author reports decode “up to 20tps with expert caching.” That’s the author’s figure, not ours.

The details from that ds4 GGUF: about 152 GiB of main weights plus 189 GiB of FP8 Engram tables that the engine reads from disk as needed, so the file has to sit on a fast local SSD. Aggressive 2-bit quantization costs some quality. Mainline llama.cpp and local Ollama don’t support the model yet. llama.cpp support is an open pull request, and Ollama offers it only as a cloud model. Unsloth’s GGUF builds run on a llama.cpp branch built from that pull request.

If you want an agent that runs fully on hardware you already own, a mid-size open model is the practical choice today. We tested one in Qwen 3.8 27B as a local agent. A middle path is Fusion: DeepSeek orchestrates from the cloud and a local model does the edits, so only the orchestrator calls the API.

The takeaway

DeepSeek V4.1 Flash 1017 is a cheap, fast model for coding agent work, and at off-peak rates it gets cheaper still. On our task it finished in under half a minute for $0.016, about 11 times less than Claude Sonnet 5.5, and its weights are open under MIT. If you use DeepSeek’s API, you already have it: deepseek-flash and the old deepseek-v4-flash both point to 1017.

Two habits keep the bill low: run batch jobs off-peak, when the price halves, and pin the provider on OpenRouter so you get the model as DeepSeek ships it. If you are still choosing the agent to run it in, our OpenCode vs Claude Code comparison and list of open-source Claude Code alternatives cover the main options.