Twenty cents. According to Deepseek’s own chart, that’s roughly what its new Flash model costs to finish a task on the DeepSWE coding benchmark. Rival models from OpenAI, Google and Anthropic land between $2 and over $10 per task on that same chart.
That’s the headline number behind Deepseek-V4.1-Flash-1017, an updated version of the Chinese AI company’s Flash model. It’s available now through Deepseek’s website, app and API. The company also published the weights as open source on Hugging Face and ModelScope, so you can download and run it yourself.
Shorter answers are where the savings come from
The model’s cost advantage mostly comes down to output length. Deepseek said the 1017 release produces shorter outputs and finishes tasks faster than its direct predecessor, V4.1-Flash, while also posting better benchmark scores. The company calls this a Pareto improvement in the trade-off between performance and cost, which is a formal way of saying it got better on one axis without getting worse on the other.
The DeepSWE chart makes the case in numbers. At its top setting, V4.1-Flash-1017 hits 77.0 percent at an average cost of just under $0.20 per task. Drop it to the low setting and it still clears 72 percent for about $0.10.
The predecessor scored 74.2 percent but cost around $0.50 per run. So the new model gains nearly three points of accuracy and costs well under half as much per task. If you’re running agents that churn through hundreds of tasks a day, that gap adds up quickly.
Bigger rivals, smaller scores (on Deepseek’s charts)
In Deepseek’s benchmarks, V4.1-Flash-1017 beats larger competing models including Claude Opus 5, GPT 5.6-Sol, Kimi K3 and GLM 5.3 in several coding and AI agent categories. On the DeepSWE cost chart, GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 all cost more per task, and none of them match the smaller Deepseek model’s accuracy.
Read that with the right amount of suspicion. These are vendor benchmarks, and independent results haven’t come out yet. Benchmarks only tell half the story anyway. Developers’ opinions of code quality can differ wildly from what a leaderboard says, and a model that scores well on a test suite can still write code you wouldn’t want to maintain.
Deepseek tuned the model for specific development environments: its own Deepseek Harness (version v0.2.1-rc.1 or later recommended), Tencent CodeBuddy and WorkBuddy. If you plug it into a different tool, don’t assume you’ll see the same numbers.
The model grades its own homework
For building complete software projects from scratch, Deepseek uses a post-training method called reinforcement learning with self-feedback, or RLSF. The model uses its own internal confidence scores or self-generated critiques as a reward signal, so there’s no need for human evaluators or an external reward model.
Deepseek said this improves end-to-end project development, from game programming to 3D modeling and design work. The model can also animate short explainer videos, similar to the recently released Opus 5.5.
Letting a model reward itself is cheap. Whether its confidence lines up with output that works is exactly the kind of thing independent testing will need to check.
Deepseek’s search engine opens up to developers
The less flashy news may matter more to anyone building agents. Deepseek is opening its web search tools, the Search & Fetch APIs, to all developers for the first time. The company said they run on a search engine it built internally that already handles hundreds of millions of queries per day worldwide.
Each endpoint costs $0.75 per 1,000 calls. Deepseek API users can use the features for free through server-side tools in Deepseek Harness, Claude Code and Codex until October 31, 2026 (4:00 PM UTC). Setup instructions are in Deepseek’s API documentation.
The catch has nothing to do with the model
Building Chinese AI models deep into your workflow got a little less appealing this week. The Chinese government reportedly plans to tighten rules on publishing open-weight models, or even block their release entirely. One deciding factor is whether a system ranks among the top 10 most capable models.
A model that beats Claude Opus 5 on coding benchmarks is exactly the kind that could get caught by that rule. If you’re thinking about building on V4.1-Flash-1017, download the weights now while they’re on Hugging Face, and keep a fallback model tested and ready in case future releases stop showing up.
