DeepSeek’s New Model Is Faster, Leaner, and Still Dirt Cheap — What V4.1-Flash-1017 Changes

Saturday afternoon, while most of the tech world was winding down for the weekend, DeepSeek quietly dropped a new AI model. No flashy keynote. No countdown timer. Just a series of posts on X and an update to its API docs.

But don’t let the quiet launch fool you. DeepSeek-V4.1-Flash-1017 — released October 17, 2026 — is the kind of update that makes expensive AI models look a little silly. It’s faster, it scores higher on coding benchmarks, and it costs a fraction of what OpenAI and Anthropic charge. Here’s what actually changed, and why you should care even if you’ve never touched an API in your life.

What DeepSeek Actually Released

The new model is called DeepSeek-V4.1-Flash-1017, and it replaces a version that only came out on September 10. That’s a refresh cycle of roughly five weeks — fast, even by AI industry standards.

DeepSeek describes the update as a “Pareto improvement” in performance and cost. Fancy term, simple meaning: things got better without anything else getting worse. Specifically, the company says the model posts stronger benchmark scores, finishes tasks faster, and produces shorter outputs than its predecessor. Developers access it through the same deepseek-flash API endpoint as before, and pricing hasn’t moved a cent.

The weights are also open. DeepSeek published them on Hugging Face under an MIT license, which means any developer can download, inspect, and build on the model freely. That’s a big deal in an era when the most capable models tend to live behind closed doors.

The Benchmark Numbers (With a Grain of Salt)

Let’s talk numbers, because DeepSeek certainly wants us to. The headline figure is a 77.0% score on DeepSWE v1.1, a coding benchmark — up from 74.2% for the previous version. For context, DeepSeek says that tops Anthropic’s Claude Opus 5 (74.0%) and OpenAI’s GPT-5.6 Sol (73.0%) on that specific test.

Other reported scores: 42.3 on Terminal-Bench 4.0 (up from 31.2), 42.6 on HLE — “Humanity’s Last Exam,” a tough reasoning test (up from 36.8) — and 29.1 on ProgramBench (up from 20.3). The company says the new checkpoint beats its predecessor on all 11 benchmarks it published, by margins of 2 to 11 points.

Here’s the honest caveat, though: these are DeepSeek’s own numbers. Independent verification hasn’t happened yet. And on raw reasoning power, the gap is real — Claude Opus 5 scores 56.3 on HLE against DeepSeek’s 42.6. So this isn’t a “we beat everyone at everything” story. It’s a “we’re shockingly good for the price” story. Which brings us to…

The Price Tag That Embarrasses Silicon Valley

This is where DeepSeek stops being just another model release and starts being a problem for its competitors.

DeepSeek charges $0.15 per million input tokens and $0.60 per million output tokens — and off-peak rates are half that. Compare that to OpenAI’s GPT-5.6 Sol at $20 per million output tokens and Anthropic’s Claude Opus 5 at $25. That’s not a small discount. That’s roughly one-thirtieth the price.

Put it in real terms: one independent pilot test found the new model completing coding tasks at about $0.0157 per task, versus $0.0577 for GPT-6.1 Sol and $0.0653 for Claude Sonnet 5.5. When you’re running thousands of tasks a day — which is exactly what AI coding assistants and agents do — that difference is the gap between a hobby project and a profitable product.

Shorter outputs sweeten the deal further. Since AI usage is typically billed by how much text the model generates, a model that reaches the same answer in fewer words is cheaper to run even when the sticker price doesn’t move. It’s a clever kind of efficiency: less rambling, lower bills.

Two Things Nobody’s Talking About Enough

1. The open Search and Fetch APIs

Buried in the same announcement: DeepSeek opened its in-house Search and Fetch APIs to outside developers for the first time. The company says its search engine indexes hundreds of billions of web pages, and access costs $0.75 per 1,000 successful requests — well under competitors like Brave ($5) and Exa ($4+) at list price. It’s a quiet expansion from “model company” to “infrastructure company,” and developers noticed: the announcement pulled thousands of likes within the hour.

2. The speed of the release cycle

Five weeks between releases. The AI race used to move in quarters; DeepSeek is moving in weeks. That pace puts real pressure on rivals who ship major updates a few times a year. Whether they can keep it up without breaking things — DeepSeek did suffer a roughly five-hour outage recently — is the open question.

What This Means for Regular People

You might never call an API directly, but this still lands in your life:

Cheaper AI apps. When the underlying model costs 3% of the alternative, developers can offer powerful features for free or cheap. Expect more capable free tiers in coding assistants, writing tools, and chatbots over the coming months.

Better open-source tools. MIT-licensed weights mean independent developers and researchers can build on this model without paying anyone or asking permission. Some of the most interesting AI tools of the next year may be built on DeepSeek’s work, not Silicon Valley’s.

Faster AI race, lower prices overall. Every time DeepSeek ships something this cheap, it drags the whole market’s pricing down. OpenAI and Anthropic have to respond — either with better models or better prices. Either way, you win.

If you’re curious how we got here, our overview of artificial intelligence breaks down the fundamentals — and if you want to see where the cutting edge is heading next, check out what quantum computing means for the future of encryption.

Frequently Asked Questions

What is DeepSeek-V4.1-Flash-1017?

It’s an updated version of DeepSeek’s lightweight AI model, released October 17, 2026. The company says it scores higher on coding and reasoning benchmarks than the September version, completes tasks faster, and produces shorter answers. Its weights are openly available under an MIT license.

How much does it cost to use?

Pricing didn’t change with this update: $0.15 per million input tokens and $0.60 per million output tokens at standard rates, with off-peak rates at half price. That makes it dramatically cheaper than comparable models from OpenAI and Anthropic.

Is it really better than Claude and GPT?

On some coding benchmarks, DeepSeek’s reported numbers beat them. On raw reasoning tests like HLE, Claude Opus 5 still leads comfortably. And remember — these are the company’s own benchmark results, not yet independently verified. Treat the claims as promising, not proven.

Can I try it myself?

Yes. The model is live on DeepSeek’s website and app, and developers can reach it through the existing deepseek-flash API endpoint. The open weights are on Hugging Face for anyone who wants to run or study the model directly.

What are the Search and Fetch APIs?

New developer tools DeepSeek opened alongside the model release. They let apps plug into DeepSeek’s own web search engine at $0.75 per 1,000 successful requests — a fraction of what competing search APIs charge.

The Bottom Line

DeepSeek-V4.1-Flash-1017 won’t end the AI race, and its boldest benchmark claims still need independent checking. But it does something arguably more important: it proves, again, that frontier-adjacent AI doesn’t have to cost a fortune. Faster, leaner, open, and cheap — that combination is why a quiet Saturday release has the whole industry talking.

The real test comes now: can DeepSeek keep shipping at this pace without the wheels coming off? If the last five weeks are any indication, we won’t have to wait long to find out.

Source: Bloomberg — “DeepSeek Pushes Out Updated AI Model It Says Is Faster, Leaner”