JOURNAL / 2026.10.18

DeepSeek V4.1 Flash 1017 improves coding and opens its web search API

The update brings MIT weights and stronger reported agent results. New search and page retrieval APIs let applications separate information gathering from the model using it, with their own costs and limits.

DeepSeek released V4.1 Flash 1017 on October 17, available on the web, in its app and through the API, with downloadable weights. The announcement combines a model update aimed at building complete projects with opening its search engine to external applications. These are related developments, but they deserve separate evaluation: writing better software and finding current documentation are different capabilities.

When we covered V4.1 Flash in September, the central change was its architecture for processing long histories with less cache. The 1017 model card now describes a minor update to that family, improving reasoning, coding, agents and vision. It retains a stated context window of one million tokens; the repository and weights use the MIT license. This is not another architecture announced from scratch.

One improvement deserves testing. DeepSeek reports a rise from 31.2 to 42.3 on Terminal-Bench 4.0, using Claude Code as the agent harness. That evaluation requires completing work in a terminal and passing each task's tests; the Artificial Analysis description explains this criterion, although its own runs use a different harness. The figures compared here are DeepSeek's, rather than an independent reproduction.

They should not be compared with the Terminal-Bench 2.1 score of 90.6 we reported in September, either. The benchmark's maintainers explain that 4.0 changes resources and tasks, requiring fresh runs. The useful signal is improvement between checkpoints within the same exam version, rather than an apparent decline between different exams. Without repeated runs and complete conditions, I would not treat the decimal as precision the evidence cannot support.

The changelog also says outputs are shorter and tasks finish faster, while the calling method and prices stay unchanged. It recommends upgrading DeepSeek Harness to 0.2.1-rc.1 or later. This is a claim about the working system: it does not establish identical savings in every application. A short response might avoid redundant steps or skip a necessary check; the final result determines which happened.

Search and retrieval move beyond the conversation

The Search and Fetch guide documents standalone JSON services: Search returns structured results, while Fetch retrieves a page as Markdown. An application can orchestrate them without asking Flash to choose every call. For example, it could retrieve recent documentation, preserve the consulted pages and hand them to another model. Downloading Flash's weights does not include the search index or its service.

This separation has concrete limits. The guide sets five requests per second per account for each service, with additional per-domain restrictions for Fetch. The published price is $0.75 per thousand successful requests for each. Server-side tool calls through the integrations specified by DeepSeek are temporarily free; billing begins October 31 at 16:00 UTC. That exception does not make all standalone searches free or remove the cost of tokens reading their results.

My assessment is that this combination deserves more attention than demonstrations of finished games. It creates the possibility of making an entire cycle cheaper: look up a dependency, implement, run tests and fix problems. It also makes it easier to produce plenty of work that looks finished. An attractive screenshot cannot tell us whether a project starts from a clean installation, preserves data or survives a later modification.

I would test the update on a small set of real assignments with acceptance criteria fixed beforehand. I would include a task whose documentation has changed and another requiring maintenance of an existing project. I would count failed attempts, web queries, review time and defects found after delivery. I would hold the harness constant when comparing models and test the new search separately. That would help distinguish better reasoning from a better information source. Convincing savings would mean less work to accept and maintain the result; a smaller token bill would be only part of that evidence.

Primary sources consulted October 18, 2026. We have not run the model or the new APIs; the cited capability results come from DeepSeek.

← Back to journal