DeepSeek V4.1 Flash Explained: The $0.15 Open AI Coding Model (2026 Guide)

On October 17, 2026, DeepSeek quietly shipped DeepSeek V4.1 Flash-1017 — and it might be the most important AI model release you barely noticed this month. The Chinese AI lab took its existing Flash model, pushed benchmark scores up across all 11 tests it reports, released the full weights under an MIT licence, and kept API pricing at a level that reads like a typo: $0.15 per million input tokens. No price hike, no fanfare, no waitlist — the update simply flowed through the existing deepseek-flash API, and the company’s web and mobile apps switched over automatically.

DeepSeek V4.1 Flash matters because it changes the economics of using frontier-class AI. On one widely watched coding benchmark, DeepSWE v1.1, the new build scores 77.0 — ahead of GPT-5.6 Sol and Claude Opus 5 on that specific test, according to DeepSeek’s own numbers — while costing a tiny fraction of what those models charge per token. Add open weights, a companion paper on memory-efficient inference, and a brand-new Search API, and you have a release that deserves a proper explainer. Here is everything DeepSeek V4.1 Flash actually is, what the numbers say, what they don’t, and whether you should switch.

DeepSeek V4.1 Flash — developer coding with the cheap open-weight AI model at night
A developer puts the new build through its paces. DeepSeek V4.1 Flash is aimed squarely at coding work, where its scores and pricing both stand out.

In this guide

What Is DeepSeek V4.1 Flash?

DeepSeek V4.1 Flash is the fast, cheap workhorse of DeepSeek’s current model lineup — the equivalent of a “budget flagship” in the phone world: most of the capability you actually use, at a fraction of the price. The “1017” in the name is simply a build tag for this October 17 checkpoint, an update to the Flash model DeepSeek first released on September 10, 2026.

DeepSeek itself describes the 1017 build as “a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks.” Minor in name, but the headline scores moved by a lot: Terminal-Bench 4.0 jumped from 31.2 to 42.3, and ProgramBench rose from 20.3 to 29.1. The company also says the model finishes tasks faster and produces shorter outputs than its predecessor — which matters because shorter outputs directly mean lower bills when you pay per token.

Three things shipped together on October 17, and it’s worth separating them:

  • The model update itself. The deepseek-flash API identifier now serves the 1017 build automatically. If you were already using Flash, you got the upgrade without changing anything — DeepSeek’s web and mobile apps moved over too. Two retired API names, deepseek-v4-flash and deepseek-v4-flash-vision-exp, now route to the same build.
  • Open weights. The full model weights were posted on Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash-1017 under the MIT licence — the most permissive licence in software. Download it, run it on your own servers, fine-tune it, even build a commercial product on it: no permission needed.
  • A Search API and a Fetch API. DeepSeek opened up its own in-house search engine — the one already powering search inside its app — as a paid developer API (more on pricing below).

One important technical note: the weights release tells us the shape of the model didn’t change. The configuration file in the Hugging Face repository is identical to the September build’s, so the architecture and the ~763-billion-parameter total are the same. This checkpoint changed the weights — what the model learned — not the model’s size or structure. Think of it as the same engine with a better tune, not a new engine.

The 1017 Update in Plain English

If all of that sounds like alphabet soup, here’s the simple version. DeepSeek V4.1 Flash is an AI model you can talk to through an API or download for free. The October 17 version is noticeably smarter than the September version at coding, reasoning and using tools — by DeepSeek’s own measurements, anyway — while costing the same as before, which was already absurdly cheap. The “Flash” name signals its design priority: speed and cost over raw power. It won’t out-think the very biggest models on the hardest questions, but for the everyday work of writing code, answering questions and running automated tasks, it now punches far above its price tag.

Why This Release Landed Now

DeepSeek V4.1 Flash arrived in the middle of an extraordinary month for efficient AI. October 2026 has seen Liquid AI ship the tiny-but-capable LFM2.5, Microsoft pre-train the massive MAI-Base-2 with an eye on efficiency, and Reflection AI release the 501-billion-parameter Beam model as open weights. The industry’s big question this autumn isn’t “who can build the biggest model” — it’s “who can deliver the most intelligence per dollar.” DeepSeek just gave the most aggressive answer yet. If you want the broader context of this efficiency race, our guides to Liquid AI’s LFM2.5 and Reflection AI Beam are worth a read alongside this one.

DeepSeek V4.1 Flash Benchmarks: The Full Numbers

Let’s start with the headline table. DeepSeek published 11 benchmark scores for the 1017 build, and every single one is higher than the September checkpoint — by 2.0 to 11.1 points. Here they are, old vs new:

DeepSeek V4.1 Flash reviewed on a laptop — faster completions at $0.15 per million tokens
Benchmarks are DeepSeek’s own measurements — impressive on their face, but not yet independently reproduced. Here’s how to read them.
Benchmark Sept 10 build DeepSeek V4.1 Flash-1017 Gain
DeepSWE v1.1 (coding) 74.2 77.0 +2.8
Terminal-Bench 4.0 (agentic terminal tasks) 31.2 42.3 +11.1
ProgramBench (programming) 20.3 29.1 +8.8
HLE without tools (hard reasoning) 36.8 42.6 +5.8
NL2Repo-Bench (code from natural language) 65.4 68.2 +2.8
Agents’ Last Exam — 35.4 new score
Apex Bench v1.1 — 72.1 new score
BabyVision with tools — 92.5 new score

The gains are genuinely large in percentage terms: Terminal-Bench rose by about 36% and ProgramBench by roughly 43%. DeepSeek also notes the model now completes tasks faster with shorter outputs — “enabling users to complete tasks faster and at lower cost,” as the model card puts it. Shorter outputs are a real, billable win: if two models charge the same per token but one needs half as many tokens to finish a job, it’s effectively half the price.

DeepSWE v1.1: 77.0 — Ahead of GPT-5.6 Sol and Claude Opus 5 (on This Test)

The eye-catching number is 77.0 on DeepSWE v1.1, a demanding benchmark for software-engineering work. For context, DeepSeek’s own figures put Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0 on that same test — so the $0.15-per-million-tokens model outscores flagship models that cost dozens of times more, here specifically. That’s the line that has the AI world doing a double-take, and it’s the comparison you’ll see most often online.

But notice the italics. This is one benchmark, measured by DeepSeek on its own test harness, and cherry-picked comparisons cut both ways. Keep that thought — we’ll come back to the other side of the ledger in a moment.

Where DeepSeek V4.1 Flash Still Trails

The same benchmark tables that make the model look heroic also show its limits, and DeepSeek published those too — to its credit. On raw reasoning power, Claude Opus 5 still leads comfortably: 56.3 vs 42.6 on Humanity’s Last Exam without tools, and 51.8 vs 42.3 on Terminal-Bench 4.0. GPT-5.6 Sol edges it on HLE as well, 44.5 to 42.6. Against Kimi K3, the new Flash wins 10 of 11 tests; against GPT-5.6 Sol, nine; against Opus 5, it trails on seven. One independent commentator summed it up bluntly: a cheap, open model with these scores is a real result, but “on par with Opus 5” is a mood, not a measurement.

The honest reading: DeepSeek V4.1 Flash is a superb value model, not the world’s smartest model. For coding tasks and everyday agentic work it is now remarkably close to the flagships; for the hardest reasoning problems, the flagships still pull away. And everything above comes from DeepSeek’s own scoreboards — no independent lab has re-run the 1017 build yet, so treat the numbers as the vendor’s best case until third parties weigh in.

Pricing: Why DeepSeek V4.1 Flash Costs Almost Nothing

DeepSeek V4.1 Flash running efficiently on data center GPU infrastructure
Cheap inference is a systems problem as much as a models problem. DeepSeek’s memory-efficient design is what lets $0.15-per-million-token pricing work.

Here’s the part that made the whole industry blink. DeepSeek did not change prices with this update — and the unchanged prices were already the lowest in the business:

  • Off-peak: $0.15 per million input tokens, $0.60 per million output tokens, $0.003 per million cached input tokens.
  • Peak hours (01:00–04:00 and 06:00–10:00 UTC, Monday to Friday): double — $0.30 input, $1.20 output.

To put that in perspective, one AI news outlet calculated that DeepSeek V4.1 Flash completes a full coding task for about $0.20, with benchmark scores rivalling models that cost over 40 times more per token — roughly one-fortieth of Claude-tier pricing. A million-token prompt — an entire novel’s worth of text — costs 15 cents off-peak. For comparison, DeepSeek’s own bigger V4 Pro charges about four times the input price and three times the output price, which makes Flash the obvious pick whenever the task doesn’t need the heavyweight model.

There are two pricing details worth knowing before you build anything on it:

  1. Peak pricing is real. If your workload runs during weekday peak windows (early-morning to mid-morning UTC), your token costs double. Schedule batch jobs off-peak and the effective price halves.
  2. Cache is nearly free. At $0.003 per million tokens, re-reading cached context costs 50 times less than fresh input. If your application re-sends long prompts — system instructions, codebases, documents — caching turns an already-cheap model into a nearly-free one.

For developers comparing the whole efficiency landscape this month — from Liquid AI’s tool-use specialist to the new wave of multi-agent coding orchestration tools — DeepSeek V4.1 Flash sets the current floor on price-per-token for genuinely capable open-weight models.

The New Search API and Fetch API

The sleeper news of October 17 wasn’t the model at all — it was the API. DeepSeek quietly opened its own in-house search engine to outside developers. This is the same search system that already powers search inside DeepSeek’s app and API, and the company says it indexes hundreds of billions of web pages.

The pricing: $0.75 per 1,000 successful requests for both the Search API and the companion Fetch API. Compare that with the established search-API market: Brave lists its search API at $5 per 1,000 requests, and Exa’s search starts at $4 per 1,000. DeepSeek is undercutting them by more than 5x. Two caveats: the features differ, so this isn’t a like-for-like comparison — and the endpoints are currently capped at 5 queries per second per account, so this is early-days capacity, not yet built for heavy production firehoses.

There’s also a small window of freebies: server-side calls routed through the DeepSeek Harness, Claude Code or Codex stay free until November 1, 2026. If you’re already building agents with those tools, the search capability lands at zero cost for the first two weeks.

Why does a search API matter in a model release? Because modern AI agents live or die by their tools. An agent that can search the live web cheaply — and a model that reasons over those results cheaply — is the combination that makes “just let the agent handle it” actually affordable. DeepSeek now sells both halves of that stack.

Open Weights (MIT): What You Can Actually Do With Them

The weights release is arguably the bigger deal for the open-source community. deepseek-ai/DeepSeek-V4.1-Flash-1017 on Hugging Face carries an MIT licence — the most permissive standard licence in software. In practice, that means:

  • Download and run it anywhere — your own servers, a cloud GPU provider, or an air-gapped machine. No API key, no account, no usage tracking.
  • Fine-tune it for your domain: medical text, legal documents, your company’s internal knowledge. The MIT licence doesn’t restrict what you train it on or what you do with the results.
  • Build commercial products on it without paying DeepSeek a cent or asking permission. That’s the real break from the open-vs-closed tradeoff — teams pricing a switch away from OpenAI or Anthropic no longer have to choose between “cheap and closed” or “open and weak.”
  • Inspect and audit it. Researchers can study exactly what changed between the September and October checkpoints, because both weight sets are public.

The practical catch is size: the weights run to about 510 GB, so self-hosting is a serious-hardware affair, not a laptop project. For most people, the API at $0.15 per million tokens is the sane route; the weights matter most to companies, researchers and fine-tuners. OpenRouter also lists the new build as deepseek/deepseek-v4-1-flash-1017 at the same $0.15/$0.60 rate, if you prefer a third-party gateway. Developers weighing open-weight options this month should also look at our Reflection AI Beam explainer — Apache 2.0 weights, a different set of tradeoffs — for the full picture of October’s open-model wave.

The KV Cache Trick Behind the Cheap Pricing

DeepSeek V4.1 Flash KV-cache compression lets 1M-token context run on less GPU memory
Memory, not raw compute, is the binding constraint on long-context inference. DeepSeek’s compression research attacks exactly that.

How can DeepSeek charge 15 cents per million tokens and survive? Part of the answer is in the companion research paper that shipped with the release: “DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression.”

Here’s the short version of the problem it solves. When a model reads a long document — say, a million tokens of context, which this model supports — it has to keep a working memory of everything it has seen so far. That working memory, called the KV cache, normally grows with every token, and storing it eats expensive GPU memory. DeepSeek’s paper is about squeezing that memory footprint dramatically: at the company’s stated figure of roughly 890 bytes per token, a full 1-million-token context needs under a gigabyte of cache in GPU memory. That is startlingly small, and it’s a big reason the per-token price can be so low.

This also explains two of the release’s headline specs. The model supports a 1-million-token context window — an entire codebase, a stack of research papers, a season of meeting transcripts in one prompt — with maximum output measured in the hundreds of thousands of tokens (DeepSeek’s documentation quotes 256K+ output tokens; coverage of the changelog cites up to 384K). And DeepSeek recommends sampling at temperature 1.0, top_p 1.0 — the defaults, in other words — suggesting the model was tuned to behave well without fiddly settings.

The memory story matters beyond this one release. Inference-side memory efficiency has been one of the busiest research threads of the past month, and whoever cracks it cheapest sets the price floor for the whole market. With this paper, DeepSeek planted its flag.

How to Try DeepSeek V4.1 Flash in 3 Steps

DeepSeek V4.1 Flash helps engineering teams ship code with fast AI coding assistance
Fast, cheap and open: the combination that makes DeepSeek V4.1 Flash a natural fit for team coding workflows.

Getting started takes minutes. Pick whichever path matches how you work:

  1. Use the API — the fastest route. Sign up at DeepSeek’s platform and call the model with the identifier deepseek-flash. That’s it — the name now points at the 1017 build, and the price didn’t change. Set temperature to 1.0 and top_p to 1.0, per DeepSeek’s recommendation. If you were already using Flash, you’re already on the new build; check your dashboard to confirm.
  2. Use the chat app. DeepSeek’s web and mobile apps switched to the 1017 build automatically with this release. Open the app you already have and you’re using it — no update or setting required.
  3. Run the weights yourself. Pull deepseek-ai/DeepSeek-V4.1-Flash-1017 from Hugging Face (MIT licence, ~510 GB download) or call it through OpenRouter as deepseek/deepseek-v4-1-flash-1017 at the same $0.15/$0.60 pricing. The Hugging Face model card documents the release in full, including the benchmark tables and the paper.

One practical tip: if you’re building anything that re-sends long context — a coding assistant that re-reads the repo, a support bot with long instructions — make sure prompt caching is on. At $0.003 per million cached tokens versus $0.15 fresh, caching is the difference between “cheap” and “basically free.”

DeepSeek V4.1 Flash vs the Competition

Here’s how the new build stacks up against the models it’s actually competing with. All rival scores below are the figures reported in the release coverage — and remember, DeepSeek’s own numbers haven’t been independently re-run yet.

Model DeepSWE v1.1 Input price / 1M Open weights?
DeepSeek V4.1 Flash-1017 77.0 $0.15 Yes (MIT)
Claude Opus 5 74.0 ~40x more No
GPT-5.6 Sol 73.0 Significantly more No
Kimi K3 Lower on 10 of 11 tests Low-cost tier Partial
DeepSeek V4 Pro (in-house) Higher on hardest reasoning ~4x Flash input Varies

The pattern is consistent: on coding-heavy benchmarks, DeepSeek V4.1 Flash now trades blows with flagships at a fortieth of the price; on the hardest general-reasoning tests, the flagships keep their lead. Note what’s not in this table: Gemini 4 Argon, which Google reports at 77.9 on DeepSWE v1.1 — a hair above Flash’s 77.0, though that figure is also vendor-reported and the model remains in limited rollout. For the full story on Google’s challenger, see our Gemini 4 Argon explainer.

Who Should Pick DeepSeek V4.1 Flash?

  • Solo developers and indie hackers: flagship-grade coding help at indie prices. This is the clearest win in the whole release.
  • Startups running agents at scale: when your agent loops hundreds of times a day, per-token price is the whole business model. Flash’s pricing plus shorter outputs is purpose-built for this.
  • Teams that need open weights: MIT licence, full weight release, fine-tune-friendly. If data can’t leave your servers, the API was never an option — the weights are.
  • Anyone who was paying flagship prices for routine coding: audit your usage. If your workload is code completion, refactoring and test-writing rather than novel research, you’re probably overpaying.

The Honest Caveats

No release this hyped deserves a free pass. Here’s what to keep in mind:

  • Every benchmark is DeepSeek’s own. The 11 scores, the comparisons with Opus 5 and GPT-5.6 Sol, the “beats Kimi K3 on 10 of 11” — all measured by DeepSeek, on harnesses DeepSeek chose. Independent reproductions haven’t landed yet. History says vendor numbers are usually directionally right and specifically optimistic; wait for third-party boards before treating 77.0 as gospel.
  • “On par with Opus 5” is chatter, not data. Some online commentary has claimed the model matches Claude Opus 5 overall. DeepSeek’s own table disagrees: Opus 5 leads on Humanity’s Last Exam (56.3 vs 42.6) and Terminal-Bench (51.8 vs 42.3). Don’t buy the hype version.
  • Peak pricing doubles the bill. $0.15 becomes $0.30 and $0.60 becomes $1.20 during weekday peak windows. Still cheap — but if you’re comparing against flat-rate competitors, use the peak number for honesty.
  • The Search API is brand new. 5 queries per second per account is a soft launch, not production scale, and “$0.75 per 1,000 successful requests” only counts successful requests — check how failures and retries are billed before you scale on it.
  • It’s a minor update, not a new generation. Same architecture, same ~763B parameter shape, new weights. The gains are real but incremental — nobody should read “1017” as “V5.”

Should You Switch? Our Verdict

DeepSeek V4.1 Flash-1017 is the best value in AI models right now, and it isn’t close. If your work is coding, agentic automation, or anything where you burn through millions of tokens a day, the combination of a 77.0 DeepSWE score, $0.15-per-million-token pricing, MIT-licensed weights and a 1M-token context is genuinely hard to argue with. The release is also a signal about where the market is going: October 2026 keeps rewarding efficiency over sheer size, and DeepSeek just raised the bar for everyone.

That said, keep two things in perspective. First, the numbers are the vendor’s until independent labs confirm them. Second, if your hardest problems are reasoning-heavy rather than code-heavy, the flagships still earn their premium — Opus 5’s lead on the toughest tests is real. For most developers, though, the smart move is simple: try the 1017 build on your actual workload this week (it costs pennies to test), measure it against whatever you’re paying for now, and let your own numbers decide. At these prices, the experiment is cheaper than the meeting about it.

FAQs

What is DeepSeek V4.1 Flash-1017?

DeepSeek V4.1 Flash-1017 is an updated checkpoint of DeepSeek’s V4.1 Flash AI model, released October 17, 2026. DeepSeek calls it a minor update with substantial improvements across reasoning, coding, agentic and visual benchmarks. The weights are free to download under the MIT licence, and the deepseek-flash API serves it at unchanged prices.

How much does DeepSeek V4.1 Flash cost?

API pricing is $0.15 per million input tokens and $0.60 per million output tokens off-peak, doubling to $0.30/$1.20 during weekday peak hours (01:00–04:00 and 06:00–10:00 UTC). Cached input costs $0.003 per million tokens. The new Search and Fetch APIs cost $0.75 per 1,000 successful requests.

Is DeepSeek V4.1 Flash really better than GPT-5.6 Sol and Claude Opus 5?

On one specific coding benchmark — DeepSWE v1.1 — DeepSeek’s own numbers put the 1017 build at 77.0, ahead of GPT-5.6 Sol (73.0) and Claude Opus 5 (74.0). On harder reasoning tests like Humanity’s Last Exam, Opus 5 still leads clearly (56.3 vs 42.6). And all of these figures are vendor-reported, not independently verified. “Better at coding per dollar” — yes, emphatically. “Better overall” — no.

Can I download and run DeepSeek V4.1 Flash myself?

Yes. The weights are on Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash-1017 under the MIT licence, which permits commercial use, fine-tuning and redistribution. The download is about 510 GB, so you’ll need serious GPU hardware — most users are better served by the API.

What is DeepSeek’s Search API?

Launched alongside the 1017 model on October 17, 2026, it’s a developer API exposing DeepSeek’s in-house web search engine (hundreds of billions of pages indexed, per the company). It costs $0.75 per 1,000 successful requests, alongside a companion Fetch API at the same price — far below Brave ($5/1K) and Exa ($4/1K) list prices, though features differ.

How does DeepSeek V4.1 Flash compare to the September build?

DeepSeek reports gains on all 11 benchmarks it tracks, from +2.0 to +11.1 points — Terminal-Bench 4.0 rose 36% (31.2 to 42.3) and ProgramBench 43% (20.3 to 29.1). The company also says tasks finish faster with shorter outputs. The architecture and model size are unchanged; the improvement comes from the new weights.

Sources: AIWeekly’s release brief (weights, Search API pricing, benchmark table), ArtificialWatch’s changelog analysis (pricing tiers, peak hours, benchmark deltas), and TECHi’s cost breakdown (KV-cache math, per-task cost estimates). All benchmark figures are DeepSeek’s own and unverified by third parties at the time of writing.

Leave a Comment

Your email address will not be published. Required fields are marked *