DeepSeek V4.1 Flash 1017: open weights, top agent scores, tiny prices

DeepSeek has a habit of shipping important things quietly. On Saturday, 17 October 2026, the Chinese lab pushed out DeepSeek V4.1 Flash 1017, a point release of the V4.1 Flash model it introduced in September. The name looks like a routine version bump, and DeepSeek itself calls it a "minor update". The numbers on the model card tell a different story. On several of the hardest agent benchmarks the jump from the September version to the October version is bigger than the gap between many competing frontier models.

The release has three parts. There is the model itself, which now runs on DeepSeek's website, app and API. There are the open weights, published under the MIT license on Hugging Face and ModelScope. And there is a new pair of developer tools, the Search and Fetch APIs, which open DeepSeek's in house search engine to everyone. In its official announcement, DeepSeek describes the model as a Pareto improvement in performance and cost: stronger benchmark results, faster task completion and shorter outputs than the previous version, all at once.

This article walks through what changed, how the model compares with closed rivals, what it costs to use, and why the release matters for anyone building with AI agents.

What the 1017 update changed

The model card on Hugging Face lists the October version next to the September one, so the progress is easy to measure. All of these scores are reported by DeepSeek itself, so they should be read as vendor claims until independent tests catch up. Still, the pattern is consistent across many benchmarks.

The biggest jump is on Terminal-Bench 4.0, a test where an AI agent has to complete real tasks inside a command line environment. The September model scored 31.2. The October model scores 42.3, a gain of more than 11 points. On ProgramBench, which asks a model to build programs that behave like a reference implementation, the score climbs from 20.3 to 29.1. On Apex Bench v1.1 it rises from 64.6 to 72.1, and on Automation-Bench from 54.8 to 60.5.

The gains are smaller but still real on the software engineering test DeepSWE v1.1, where the score goes from 74.2 to 77.0, and on Agents' Last Exam, from 31.8 to 35.4. Even the general reasoning test Humanity's Last Exam without tools improves from 36.8 to 42.6.

DeepSeek says the improvements come from reinforcement learning with self feedback aimed at end to end project development. The announcement shows the model building complete projects from a single request: a 3D model of the Forbidden City, a browser racing game, a recreation of Portal 2 in the Godot engine and an animated scene that runs from prehistoric times to the modern era. Those demos are marketing, of course, but they match the benchmark story. This update is about agents that finish whole jobs, not about a chatbot that writes nicer paragraphs.

There is another detail that matters for cost. According to DeepSeek, the new version completes tasks faster and produces shorter outputs. For an agent that runs for many steps, shorter outputs mean fewer billed tokens per finished task, even before any change in price.

How it compares with the closed frontier

The more surprising part of the model card is the comparison with closed models. DeepSeek lists its new release next to GPT-5.6 Sol from OpenAI, Opus 5 from Anthropic, K3 from Moonshot and GLM-5.3 from Zhipu.

On DeepSWE v1.1 the open model scores 77.0, ahead of Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. On Automation-Bench it posts 60.5, compared with 50.3 for Opus 5 and 45.8 for GPT-5.6 Sol. On Agents' Last Exam it reaches 35.4, again ahead of every model in the table. On Apex Bench v1.1 it leads with 72.1 against 65.8 for Opus 5.

It does not win everywhere, and the card is honest about that. On Terminal-Bench 4.0, Opus 5 still leads clearly with 51.8, while DeepSeek reaches 42.3 and GPT-5.6 Sol 39.9. On NL2Repo-Bench, which asks a model to build an entire repository from a written description, Opus 5 scores 75.3 against DeepSeek's 68.2. On ProgramBench, Opus 5 leads with 37.0 against 29.1. And on Humanity's Last Exam without tools, Opus 5 sits far ahead at 56.3, while DeepSeek reaches 42.6 and GPT-5.6 Sol 44.5. The visual benchmarks show the same picture: close to the leaders, but usually just behind Opus 5.

Two caveats are worth keeping in mind. First, the test setup matters. DeepSeek ran Terminal-Bench 4.0 inside the Claude Code harness and DeepSWE v1.1 inside the mini SWE v2 harness, and the original V4.1 Flash card shows that scores move noticeably from one agent framework to another. Second, these are numbers chosen and measured by the company that built the model. Independent leaderboards will need a few weeks to confirm or correct them.

Even with those caveats, the headline holds. A model whose weights anyone can download is now trading blows with the best closed systems on practical agent work. That was not true a year ago.

Why the price is the real story

Benchmarks get the attention, but pricing decides who actually uses a model. DeepSeek's pricing page now lists DeepSeek V4.1 Flash 1017 as the model behind the `deepseek-flash` name, and the numbers are very low.

Outside peak hours, one million input tokens cost $0.15 when they are not cached and just $0.003 when they hit the cache. One million output tokens cost $0.60. During peak hours every price doubles, to $0.30, $0.006 and $1.20. DeepSeek defines peak hours as 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. All other hours, including entire weekends and Chinese public holidays, are billed at the off peak rate.

For comparison, DeepSeek's own larger V4 Pro model costs $0.66 per million uncached input tokens and $1.98 per million output tokens off peak. On several agent benchmarks the cheaper Flash 1017 model now scores higher than V4 Pro did in the September comparison table, which makes the bigger model hard to justify for many coding workloads.

The cache price deserves special attention. Coding agents send the same large context again and again: the repository, the instructions, the history of the session. When most of that context is cached, the effective input cost falls close to zero. Combined with shorter outputs, this is where DeepSeek's claim of cheaper finished tasks comes from.

There is also a practical change for existing customers. The older model names `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` still work, but DeepSeek says those models have been retired. Requests sent to them are now served by V4.1 Flash 1017 and billed at the Flash price. Developers who never touched their code are already using the new model.

Search and Fetch: DeepSeek opens its search engine

The second announcement may matter just as much for agent builders. DeepSeek has opened two new endpoints, Search and Fetch, built on what it calls its in house search engine, which it says already handles hundreds of millions of requests a day. Search returns web results for a query. Fetch retrieves the content of a page. Each costs $0.75 per 1,000 successful calls.

There is a launch offer, too. Search and Fetch calls made through server side tools inside DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users for now. According to the announcement, billing for those calls starts on 31 October 2026 at 16:00 UTC.

This matters because web access is one of the hidden costs of running agents. An agent that researches a bug, reads documentation or checks a library version may call a search tool dozens of times per task. Having a cheap search and fetch layer from the same provider as the model removes one more vendor from the stack. It also shows that DeepSeek wants to compete for the whole agent workflow, not only for the model call in the middle.

The list of supported tools is telling as well. DeepSeek's documentation now includes integration guides for Claude Code, Codex, OpenCode and its own DeepSeek Harness. The company optimised the release for DeepSeek Harness and for Tencent's CodeBuddy and WorkBuddy, and recommends upgrading DeepSeek Harness to version 0.2.1 rc.1 or later for the best results.

What is inside the model

V4.1 Flash 1017 builds on the architecture DeepSeek described for V4.1 Flash in September. That model card explains why a large model can be this cheap to run.

The model is a Mixture of Experts design with 552 billion backbone parameters. Instead of running all of them for every token, it routes each token to a small set of specialist sub networks. Each MoE layer has one shared expert and 384 routed experts, and six routed experts are active per token. The result is that only about 8 billion parameters are active per token while reading input and about 16 billion while generating output. That split is deliberate: agent workloads read far more than they write, so making the reading phase cheap saves the most money.

The second trick is memory. Long context is expensive mainly because of the KV cache, the stored attention data the model needs for every token it has already seen. DeepSeek says V4.1 Flash needs only 890 bytes of global KV cache per token, about a quarter of what V4 Flash needed. That is what makes the one million token context window affordable, and the API allows up to 384,000 tokens of output.

The model is also natively multimodal. It was trained from scratch on 45 trillion tokens of text and images, with its own vision encoder, so the same model can read screenshots, charts and documents. DeepSeek's reasoning effort setting can be adjusted on a scale from 1 to 100, which lets developers trade accuracy for cost per request.

The weights come in FP8 format and are released under the MIT license, one of the most permissive licenses in software. Companies can run, modify and fine tune the model commercially without asking DeepSeek for permission. Running the full model locally is still a job for serious server hardware, not a laptop, but cloud providers and inference startups can host it right away. Within a day of release, the Hugging Face page showed more than a thousand likes and more than ten thousand downloads.

What it means

Three things stand out.

First, the gap between open and closed models on agent work is now small, at least according to the vendor's own numbers. If independent tests confirm even part of these results, teams that pay premium prices for coding agents will have a serious reason to test an open alternative.

Second, the price pressure on everyone else keeps growing. A model that matches or beats GPT-5.6 Sol on several agent tests while charging $0.60 per million output tokens off peak forces other labs to defend their prices with better quality, better tools or better trust.

Third, DeepSeek is building a platform, not just shipping models. Free search inside popular coding tools, documented support for Claude Code and Codex, and automatic upgrades for customers on the old model names are all moves designed to make switching easy.

The open questions are the usual ones. How well do these scores hold up outside DeepSeek's own test setups? How does the model behave on long, messy real world projects? And how will companies weigh data residency and compliance concerns when they use a Chinese provider's hosted API? The open weights help with the last point, because organisations can host the model themselves.

For now, DeepSeek V4.1 Flash 1017 is another reminder that a "minor update" from this lab can move the whole market. Coverage of the launch is also available from The Decoder.

Sources: DeepSeek announcement, DeepSeek V4.1 Flash 1017 model card, DeepSeek V4.1 Flash model card, DeepSeek API pricing, ModelScope release, The Decoder.