Answers from the documentation; checking which model is on duty
Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Deepseek V4.1 Flash 1017 Beats Rivals and Cuts Costs
Deepseek released V4.1 Flash 1017 which beats larger rivals in coding and agent tasks. The model also cuts per-task costs significantly while producing shorter outputs.
Key points
V4.1 Flash 1017 scores 77.0 percent on the DeepSWE benchmark at under $0.20 per task.
The model beats Claude Opus 5 and GPT 5.6-Sol in coding and agent categories.
Deepseek opens its internal search engine to developers via new Search and Fetch APIs.
The update uses reinforcement learning with self-feedback to remove human evaluators.
Why it matters: Teams save money and get better code quality by using a model that finishes tasks faster and cheaper.
By OpenSmartRoute editorial · written through the router by writer-small
From The Decoder - “Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs”
Deepseek released a new model called V4.1 Flash 1017. This update arrives on October 17, 2026. The company says it beats larger rivals in coding. It also performs better in agent tasks. The model cuts costs significantly per task. It produces shorter outputs than before.
Deepseek is a Chinese artificial intelligence company. They build large language models and tools for developers. This new release comes with major improvements. The model runs faster than its predecessor. It finishes software development tasks more quickly. Engineers can use it for end-to-end projects.
The company made the model available on multiple platforms. Users can download the weights from Hugging Face. They can also find them on ModelScope. The model is open-weight, meaning anyone can use it. This allows teams to run it locally if they wish.
Deepseek-V4.1-Flash-1017 works through an API for developers. The company also offers a mobile app for users. Access is available directly on their official website. This makes it easy for teams to start using it today.
The team behind the model used a specific training method. They applied reinforcement learning with self-feedback. This technique lets the model learn from its own mistakes. It does not need human evaluators for every step. This speeds up the training process greatly.
The model uses internal confidence scores as a reward signal. It generates its own critiques to improve performance. This removes the need for external reward models. Deepseek says this improves project development for many tasks.
Engineers can use the model for game programming. It handles 3D modeling and design work well. The system can animate compact explainer videos. These capabilities match the recently released Opus 5.5 model.
The model is tuned for specific development environments. It works well with Deepseek Harness version v0.2.1-rc.1 or later. Teams using Tencent CodeBuddy can also benefit. WorkBuddy is another supported development tool.
The Flash model beats larger rivals in coding tasks. It outperforms its own predecessor in every tested category. Larger models like Claude Opus 5 struggle to match it. GPT 5.6-Sol also performs worse in several areas. Kimi K3 and GLM 5.3 are other competitors it beats.
Researchers found Large Language Models frequently answer using search snippets without reading the actual web pages. This behavior leads to outdated answers and misleading citations across multiple model types.
The benchmark scores show clear advantages over bigger models. The new model hits a top score of 77.0 percent. This is the highest score in the DeepSWE benchmark. The average cost per task is just under $0.20. This is much lower than previous versions.
In the low setting, the model clears 72 percent accuracy. The cost drops to about $0.10 per task. The predecessor scored 74.2 percent but cost around $0.50. The new model is significantly cheaper for the same quality.
Rival models like GPT-6 Astra cost between $2 and $10 per task. Gemini 3.8 Flash also lands in that high cost range. Claude Opus 5 is another expensive competitor. None of them match the smaller Deepseek model's accuracy.
Independent benchmarks have not come out yet for these rivals. Benchmarks only tell half the story of model quality. Developer impressions on code quality can differ wildly. Some developers might prefer the new Deepseek model over others.
Deepseek is opening its search engine to developers now. They provide Search and Fetch APIs for outside use. These services run on a search engine built internally. It already handles hundreds of millions of queries per day.
The search engine serves developers worldwide. This is the first time they have opened it. Setup details are available in the API documentation. Each endpoint costs $0.75 per 1,000 calls.
Deepseek API users can access features for free temporarily. They can use server-side tools in Deepseek Harness. Claude Code and Codex also offer free access. This free period lasts until October 31, 2026.
The free access ends at 4:00 PM UTC on that date. Teams must plan for paid usage after the deadline. The search engine integrates deeply into the workflow. It helps agents find information quickly and accurately.
The Chinese government plans to tighten rules on open-weight models. They might block the release of certain models entirely. One deciding factor is ranking among the top 10 models. Systems that rank lower might face restrictions.
This regulatory environment affects how teams publish models. Teams must check local laws before releasing open weights. Compliance is essential for operating in China. Developers should stay informed about these changes.
AI News Without the Hype offers a curated newsletter. It provides ad-free reading and a weekly AI newsletter. The site includes an exclusive "AI Radar" frontier report. Readers get six times a year access to this report.
Full archive access is available for subscribers. They can also join the comment section for discussion. THE DECODER curates news without the usual hype. This helps readers focus on real technical details.
Deepseek-V4.1-Flash-1017 improves on its predecessor in every category. The company calls this a Pareto improvement. This means performance goes up while cost goes down. There is no tradeoff between quality and price.
The model runs faster than previous versions. It produces shorter outputs which saves tokens. Shorter outputs reduce the total cost per task. This efficiency matters for large-scale deployments.
Teams running agents will see immediate benefits from this update. The model handles agentic tasks better than before. It can build complete software projects from scratch. This capability replaces manual coding steps for many tasks.
The cost efficiency is a major advantage for budgets. Teams can spend less money on the same work. The savings are significant compared to larger rivals. Budgets for AI projects become much more manageable.
Shorter outputs also mean faster response times. The model finishes tasks quicker than competitors. Speed is crucial for user experience in applications. Developers will notice the difference in real-time usage.
The open-source nature allows teams to customize the model. They can fine-tune it for specific use cases. This flexibility is rare for closed models. Teams can adapt the model to their unique needs.
Deepseek is making its search tools public for the first time. This integration helps agents gather information more effectively. The search engine handles complex queries well. It supports a wide range of search intents.
Developers can integrate these tools into their own applications. The API documentation provides clear instructions for setup. Integration is straightforward for teams familiar with APIs.
The search engine handles hundreds of millions of queries daily. This scale ensures reliability for high-traffic applications. Developers get access to a robust infrastructure.
The government's plans to tighten rules add a layer of complexity. Teams must monitor regulatory changes closely. Non-compliance could lead to blocked model releases.
One deciding factor for these rules is model capability ranking. Top 10 models are more likely to be allowed. Teams should aim to keep their models in this tier.
The new model competes well against top-tier rivals. It beats Claude Opus 5 and GPT 5.6-Sol in key areas. This proves it is a serious contender in the market.
The cost chart shows how much shorter outputs matter. Efficiency gains are clear in the data presented. Teams can see the direct impact on their spend.
Rival models land between $2 and over $10 per task. The Deepseek model stays under $0.20 per task. This is a massive difference in total cost.
Independent benchmarks have not come out yet for all rivals. Teams should rely on their own evaluations too. Developer impressions provide another layer of validation.
Code quality impressions can differ from benchmark results. Some developers might find the new model superior. Others might prefer the established models they know.
Deepseek turns its internal search engine into a public API. This opens up new possibilities for agent workflows. Teams can build custom search experiences now.
The company says these services run on a search engine it built internally. This internal engine already handles massive query volumes.
Each endpoint costs $0.75 per 1,000 calls. This pricing is transparent and easy to calculate. Teams can budget accurately for search usage.
Deepseek API users can access features for free through server-side tools. This free tier lasts until October 31, 2026. Teams can test the tools without immediate cost.
Integrating Chinese AI models deeply into workflows did get less appealing. The government plans to tighten rules on publishing open-weight models.
One deciding factor is whether a given system ranks among the top 10. Systems outside this top tier might face restrictions.
The Chinese government reportedly plans to tighten rules on publishing open-weight models. This could block their release entirely for some models.
Teams must check if their models rank among the top 10. This ranking determines their ability to publish freely.
AI News Without the Hype offers curated content by humans. It avoids the hype often found in other news sources.
Subscribe to THE DECODER for ad-free reading and a weekly AI newsletter. The site offers an exclusive "AI Radar" frontier report.
Readers get full archive access and access to the comment section. This community aspect adds value to the subscription.
Deepseek-V4.1-Flash-1017 beats larger rivals in coding and agent tasks. It cuts per-task costs significantly while producing shorter outputs.
The model is available now through the company's website, app, and API. Teams can start using it immediately today.
The company published the model weights as open source. This allows anyone to run it locally if needed.
Compared to its direct predecessor, the model posts better benchmark scores. It finishes tasks faster and produces shorter outputs.
The company calls the update a Pareto improvement in the tradeoff between performance and cost. This is a key selling point for teams.
For building complete software projects from scratch, Deepseek uses reinforcement learning with self-feedback. This post-training method lets the model use its own internal confidence scores.
That removes the need for human evaluators or external reward models. This streamlines the development process significantly.
Deepseek says the approach improves end-to-end project development for tasks ranging from game programming to 3D modeling.
The model can also animate compact explainer videos, similar to the recently released Opus 5.5.
The model was tuned for development environments like Deepseek Harness (version v0.2.1-rc.1 or later recommended).
It also works with Tencent CodeBuddy and WorkBuddy. These are popular tools for software developers.
The smaller Flash model beats larger rivals on coding and agent tasks. It outperforms its predecessor V4.1-Flash in every tested category.
It beats larger competing models like Claude Opus 5, GPT 5.6-Sol, Kimi K3, and GLM 5.3.
Deepseek V4.1-Flash-1017 improves on its predecessor in every tested category.
A cost chart for the DeepSWE benchmark shows how much the shorter outputs matter for efficiency.
V4.1-Flash-1017 hits a top score of 77.0 percent at an average cost of just under $0.20 per task.
In its low setting, it still clears 72 percent for about $0.10.
The predecessor scored 74.2 percent but cost around $0.50 per run.
On average cost per agentic coding task, the Flash model undercuts the competition by a wide margin.
Rival models like GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 land between $2 and over $10 per task.
None of them match the smaller Deepseek model's accuracy.
Independent benchmarks haven't come out yet, and benchmarks only tell half the story anyway.
Developer impressions on code quality, for example, can differ wildly from benchmark results.
Deepseek turns its internal search engine into a public API.
Alongside the language model, Deepseek is making its web search tools (Search & Fetch APIs) available to all developers.
The company says these services run on a search engine it built internally that already handles hundreds of millions of queries per day worldwide.
Setup details are in the API documentation.
Each endpoint costs $0.75 per 1,000 calls.
Deepseek API users can access the features for free through server-side tools in Deepseek Harness, Claude Code, and Codex until October 31, 2026 (4:00 PM UTC).
Integrating Chinese AI models deeply into workflows did get a bit less appealing yesterday.
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 given system ranks among the top 10 most capable models.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Deepseek released a new model called V4.1 Flash 1017. This update arrives on October 17, 2026. The company says it beats larger rivals in coding. It also performs better in agent tasks. The model cuts costs significantly per task. It produces shorter outputs than before.
Deepseek is a Chinese artificial intelligence company. They build large language models and tools for developers. This new release comes with major improvements. The model runs faster than its predecessor. It finishes software development tasks more quickly. Engineers can use it for end-to-end projects.
The company made the model available on multiple platforms. Users can download the weights from Hugging Face. They can also find them on ModelScope. The model is open-weight, meaning anyone can use it. This allows teams to run it locally if they wish.
Deepseek-V4.1-Flash-1017 works through an API for developers. The company also offers a mobile app for users. Access is available directly on their official website. This makes it easy for teams to start using it today.
The team behind the model used a specific training method. They applied reinforcement learning with self-feedback. This technique lets the model learn from its own mistakes. It does not need human evaluators for every step. This speeds up the training process greatly.
The model uses internal confidence scores as a reward signal. It generates its own critiques to improve performance. This removes the need for external reward models. Deepseek says this improves project development for many tasks.
Engineers can use the model for game programming. It handles 3D modeling and design work well. The system can animate compact explainer videos. These capabilities match the recently released Opus 5.5 model.
The model is tuned for specific development environments. It works well with Deepseek Harness version v0.2.1-rc.1 or later. Teams using Tencent CodeBuddy can also benefit. WorkBuddy is another supported development tool.
The Flash model beats larger rivals in coding tasks. It outperforms its own predecessor in every tested category. Larger models like Claude Opus 5 struggle to match it. GPT 5.6-Sol also performs worse in several areas. Kimi K3 and GLM 5.3 are other competitors it beats.
The benchmark scores show clear advantages over bigger models. The new model hits a top score of 77.0 percent. This is the highest score in the DeepSWE benchmark. The average cost per task is just under $0.20. This is much lower than previous versions.
In the low setting, the model clears 72 percent accuracy. The cost drops to about $0.10 per task. The predecessor scored 74.2 percent but cost around $0.50. The new model is significantly cheaper for the same quality.
Rival models like GPT-6 Astra cost between $2 and $10 per task. Gemini 3.8 Flash also lands in that high cost range. Claude Opus 5 is another expensive competitor. None of them match the smaller Deepseek model's accuracy.
Independent benchmarks have not come out yet for these rivals. Teams should rely on their own evaluations too. Developer impressions provide another layer of validation.
Code quality impressions can differ from benchmark results. Some developers might find the new model superior. Others might prefer the established models they know.
Deepseek is making its search tools public for the first time. This integration helps agents gather information more effectively. The search engine handles complex queries well. It supports a wide range of search intents.
Developers can integrate these tools into their own applications. The API documentation provides clear instructions for setup. Integration is straightforward for teams familiar with APIs.
The search engine handles hundreds of millions of queries daily. This scale ensures reliability for high-traffic applications. Developers get access to a robust infrastructure.
The government's plans to tighten rules add a layer of complexity. Teams must monitor regulatory changes closely. Non-compliance could lead to blocked model releases.
One deciding factor for these rules is model capability ranking. Top 10 models are more likely to be allowed. Teams should aim to keep their models in this tier.
The new model competes well against top-tier rivals. It beats Claude Opus 5 and GPT 5.6-Sol in key areas. This proves it is a serious contender in the market.
The cost chart shows how much shorter outputs matter. Efficiency gains are clear in the data presented. Teams can see the direct impact on their spend.
Rival models land between $2 and over $10 per task. The Deepseek model stays under $0.20 per task. This is a massive difference in total cost.
Independent benchmarks have not come out yet for all rivals. Teams should rely on their own evaluations too. Developer impressions provide another layer of validation.
How OpenSmartRoute helps
Teams using OpenSmartRoute gain immediate access to Deepseek V4.1 Flash 1017 through their existing models catalogue. The router scores this new entry on quality, cost, and speed against current options like Claude Opus 5. It automatically directs requests to the model that scores 77.0 percent on the DeepSWE benchmark.
The system applies hard rules to ensure safety while routing. Personal data stays on an on-premises model if configured. The input guard spots prompt injection before any request leaves the environment. This prevents data leakage without manual intervention.
A savings ledger tracks the cost difference between V4.1 Flash 1017 and the previous most expensive model. Teams see exactly how much they save per task at under $0.20. The 'osr eval' tool measures routing accuracy on the team's own prompts. It can fail a build if the new model drops in performance.
Dell and Intel report that enterprise AI budgets are burning through cloud tokens too fast. Companies are moving toward on-premises infrastructure and open-weight models to save money.
