Deepseek has released a new open-weight model with version V4.1-Flash-1017. Open weight means: the trained model weights are freely available, so the model can also be used outside the provider's infrastructure. Compared with the predecessor, Deepseek reports better benchmark results, faster processing of tasks and shorter answers.
The last point in particular is more relevant for productive use than it initially sounds. Shorter outputs mean fewer generated tokens and therefore usually lower costs per request as well as shorter waiting times in the editor.
Search and fetch APIs as new building blocks
Two interfaces that connect the model with external information are new: a search API for search queries and a fetch API for retrieving content. According to Deepseek, both cost 0.75 US dollars per 1,000 requests. For API users in tools such as Claude Code and Codex, they are free for the time being.
The provider thus addresses a typical gap in AI-supported development workflows: a language model only knows the state of its training data. It must actively obtain documentation on current framework versions, changelogs or error messages from the web. Standardised search and retrieval interfaces take the task of building and maintaining this connection themselves off teams' hands.
Why compatibility with coding tools matters
Claude Code and Codex are command-line and editor tools that delegate development tasks to a language model: reading code, changing it, running tests, preparing refactorings. What matters is that such tools increasingly allow interchangeable models behind them. The interface and the workflow remain the same, the model underneath can be swapped.
That is exactly what makes a cheaper, openly available model interesting. Anyone who has so far paid premium prices for every request in everyday development can shift parts of the load – such as routine tasks, code explanations or first drafts – and reserve more expensive models for complex architecture questions.
Cost structure instead of unit price
In practice, the costs of AI-supported development rarely arise from individual spectacular requests, but from volume: many small interactions across many developers, over months. A model that solves the same task with fewer tokens and in less time therefore has a double effect – through the price per token and through the quantity of generated tokens.
For the new interfaces, a sober calculation is worthwhile. 0.75 US dollars per 1,000 requests sounds low, but adds up when automated agents search and fetch multiple times per task. The currently free use in Claude Code and Codex is marked as provisional; a calculation should be based on the regular price.
Independence from individual providers
The strategically more interesting aspect is the availability of the weights. Open-weight models can in principle also be operated on own or European infrastructure. For organisations with requirements regarding data storage, compliance or resilience, this is a different starting point than a pure cloud API whose terms and availability are determined solely by the provider.
At the same time, open weight does not mean simple self-hosting. Operating larger models requires suitable hardware, operational knowledge and ongoing maintenance. For many medium-sized companies, the API route remains the more pragmatic path – but the option to switch increases their negotiating position and reduces the risk of being tied to a single platform.
What can be checked before use
- Test your own task classes: Benchmark values say little about the specific codebase. A comparison on real tickets, legacy code and typical review tasks is more useful.
- Evaluate shorter outputs: More concise answers save costs but can shorten justifications and context. Whether that is acceptable in the team only becomes clear in everyday use.
- Prepare for model changes: Keep prompts, configurations and agent rules in such a way that swapping the model does not become a rebuild of the entire workflow.
- Clarify data flows: Which code excerpts and internal information leave the company when an external model or a search API is used?
- Set cost ceilings: Budgets and limits per project prevent automated agents from generating high request volumes unnoticed.
Classification in a fast market
The intervals between model releases are getting shorter, and with each step the ratio of performance to price shifts. For project managers, this means the question is less which model is the best today, but how their own development environment can be set up so that a change remains possible without great effort.
Model-independent tools, documented prompts and clear interfaces are more important here than the choice of a specific provider. They turn an individual decision into a revisable option.
For web and software projects, this means concretely: cheaper, open models make AI support economically viable even in budgets where it was previously difficult to justify – for example in maintenance, migrations or documentation. What matters is keeping your own toolchain interchangeable instead of optimising it for a single model.
And regardless of price, responsibility for quality, safety and review processes remains with the team, not the model.