DeepSeek V4.1-Flash-1017: Faster, Cheaper AI Agents and a Search API
· Cortex Solutions
On October 17, 2026 DeepSeek updated its Flash model to V4.1-Flash-1017 with higher agent and coding scores, shorter outputs, and opened its search engine as an API.
On Saturday, October 17, 2026, Chinese AI company DeepSeek released an update to its fast model, now called DeepSeek-V4.1-Flash-1017. According to the official DeepSeek API changelog dated the same day, the new version replaces V4.1 Flash on the DeepSeek website, app and API, and developers keep calling it with the same model name, deepseek-flash, with no change to the calling method or the price. Bloomberg also covered the launch the same day.
What changed in the model?
DeepSeek describes the update as stronger on benchmarks, faster at completing tasks and producing shorter answers than the previous version. The tech outlet The Decoder, reporting on October 17, quoted the company calling it a "Pareto improvement" in the trade-off between performance and cost: better results without paying more.
The changelog lists the model's scores on a set of benchmarks focused on programming and AI agents, including:
- DeepSWE v1.1 (software engineering tasks): 77.0
- Terminal-Bench 4.0 (working in a command-line terminal): 42.3
- NL2Repo-Bench (building a code repository from a description): 68.2
- Automation-Bench v1.0.6: 60.5
- Humanity's Last Exam without tools: 42.6
DeepSeek says it reduced task runtime and output length, which means tasks finish faster and cost less. It tuned the model to work with its own agent environment, DeepSeek Harness, and with Tencent's CodeBuddy and WorkBuddy tools.
Cost per task: the headline number
According to The Decoder, at its highest setting the model scored 77.0% on DeepSWE at an average cost of just under $0.20 per task, and above 72% at around $0.10 per task on a lower setting, compared with 74.2% at around $0.50 per run for the previous V4.1-Flash. The outlet adds that the competing models shown in DeepSeek's cost chart cost roughly $2 to more than $10 per task.
The Decoder also reports that the model's weights were published openly on Hugging Face and ModelScope, and that post-training relied on reinforcement learning with self-feedback, where the model's own confidence and critiques act as the reward signal instead of human reviewers.
A search engine for developers
Alongside the model, DeepSeek announced in the same changelog that it is opening its in-house search engine, which powers search on its website and API. The company describes it as an AI-native general-purpose search engine that indexes hundreds of billions of pages, supports multiple languages and filters SEO spam and low-quality AI-generated content. The Decoder says the Search and Fetch APIs are priced at $0.75 per 1,000 calls per endpoint, with free access until October 31, 2026 for API users through certain agent tools.
Important caveats
All of these results come from DeepSeek's own tests. The Decoder notes that independent benchmarks had not yet been published, and that developers' impressions of code quality can differ from benchmark numbers. It also mentions reports that Chinese authorities may tighten rules on publishing open-weight models. Companies should therefore test the model on their own tasks before relying on it.
What does this mean for Saudi businesses?
- The cost of running AI agents keeps falling: when a programming or automation task costs cents instead of dollars, automating document processing, customer replies and internal reports becomes viable for small and medium businesses, not only large enterprises.
- Open weights mean more options for hosting: a company can consider running a model on its own servers or with a local cloud provider, which matters when handling customer data and complying with the Personal Data Protection Law and the data classification rules in the Kingdom.
- Do not build on one model only: models change every few weeks. It is smarter to design your systems so the model can be swapped without rebuilding everything.
- Evaluate before you adopt: run a small pilot on real, non-sensitive data in Arabic and English and measure quality, speed and cost yourself before relying on any vendor's benchmark claims.
At Cortex Solutions we help companies design practical AI agents and automation connected to their systems, choosing the right model for each task and keeping data under control. Learn more: https://cortex.sa/services/ai-automation
Sources: the DeepSeek API changelog (October 17, 2026), The Decoder (October 17, 2026), and Bloomberg (October 17, 2026).