DeepSeek updates V4.1-Flash and opens Search and Fetch APIs

A new checkpoint, not the first V4.1-Flash release

DeepSeek has released V4.1-Flash-1017, an update to its Flash model. The API documentation identifies it as the active version behind the deepseek-flash alias. It is not the initial V4.1-Flash launch, which was introduced in September: the 1017 suffix distinguishes this later checkpoint, announced on October 17, 2026.

The company highlights improvements in coding and agentic tasks, along with shorter responses and lower costs in its own evaluation. Those figures are results reported by DeepSeek, not an independently replicated comparison of the new model. The distinction matters: a strong benchmark score can help identify a candidate for testing, but it does not guarantee the same outcome with different prompts, tools, or real-world tasks.

There is also a separate product change. Search and Fetch are now available as standalone APIs, allowing applications and agents to search the web and retrieve pages without relying solely on an internal tool. The documentation confirms their prices per successful request and specifies a temporary exception for calls made through certain agent environments.

DeepSWE: a reported improvement, with limits on interpretation

For the DeepSWE v1.1 test, DeepSeek reports a score of 77.0 for V4.1-Flash-1017 at maximum reasoning effort, with an average cost of less than US$0.20 per task. As a reference point, the previous V4.1-Flash checkpoint scored 74.2 in an earlier technical table and cost about US$0.50 per run in the chart cited around the launch. These are DeepSeek evaluation results; they are not equivalent to an independent, controlled test of the new model.

The agent scaffold itself can substantially affect the result. In its earlier DeepSWE v1.1 evaluation, DeepSeek reported different scores for the same model when the environment and tool-execution system changed. For version 1017, the available extract does not provide a complete methodology that would allow every factor—such as the number of attempts, task selection, or infrastructure cost—to be independently compared.

The 77.0 score should not be turned into a blanket claim that the model outperforms every rival. DeepSeek’s published comparisons include models from other labs, but they were run within its own evaluation setup. Until the 1017 checkpoint is independently tested under equivalent conditions, the sensible approach is to treat the results as a signal for choosing candidates to evaluate, not as a definitive ranking.

Search and Fetch: two APIs billed per request

DeepSeek offers Search to find information on the web and Fetch to retrieve content from pages. According to the official pricing table, each API costs US$0.75 per 1,000 successful requests—about US$0.00075 per completed call. Billing is based on requests, not on the number of tokens generated by the model.

The documentation also sets out a temporary exception: calls made through DeepSeek Harness, Claude Code, or Codex will not incur per-request charges until billing begins at 00:00 on November 1, 2026, Beijing time. That does not mean access is free through every route, or that the free period continues beyond that date. Anyone integrating the endpoints directly should budget using the published Search and Fetch rates.

For an agent, search and retrieval serve different steps. Search finds relevant pages; Fetch obtains content from a specific page. Keeping them separate makes it easier to track how many calls are made and where the cost arises. The documentation available in the sources here confirms prices and availability, but does not provide details such as quotas, rate limits, or every query option. Check the current guide before designing a production integration.

Available weights and the MIT License

DeepSeek’s official Hugging Face repository publishes the V4.1-Flash weights under the MIT License. The model’s technical materials also describe a multimodal mixture-of-experts architecture with 552 billion backbone parameters, a context window of up to one million tokens, text and image input, and text output. The license permits broad use, modification, and redistribution, subject to its terms.

The availability of weights does not mean the model is small or easy to run. The architecture activates 8 billion parameters per token during input processing and 16 billion during generation, but its total parameter count and memory requirements remain substantial. For many teams, a hosted API will be more practical than managing hardware, weight conversion, and inference serving.

It is also important to distinguish open weights from a complete release of the training process. The repository provides the weights and the technical materials described, but that does not establish that all training data, filters, infrastructure, or development decisions are available. The MIT License sets out permissions for the covered artifacts; it does not remove the need to review privacy, security, local obligations, or the terms of external services.

What to measure before adopting it

An improvement in an evaluation is not, on its own, a reason to migrate an entire agent. Measure answer quality, actual tool completion, latency, and cost per accepted result separately. A model may generate fewer tokens and still cost more if it requires extra retries, fails more often, or needs an additional validation step.

For software tasks, measure whether changes compile and pass tests, not merely whether the responses sound correct. For search agents, check whether retrieved sources answer the query and whether the system cites or preserves the necessary evidence. In both cases, retain human approval for irreversible actions, production changes, or security decisions.

The available documentation confirms that the direct API uses the deepseek-flash alias and that Search and Fetch have separate rates. The full operational details for these endpoints—for example, quotas, formats, and limits—are not included in the available extract, so they should not be guessed. Before deployment, review the current official guide and pin the version in your tests to avoid attributing future changes to the wrong checkpoint.

Bottom line: a promising update that still needs outside validation

V4.1-Flash-1017 combines an update to the Flash model with the release of Search and Fetch as separate services. It may interest developers building agents that alternate between reasoning, coding, and web search, particularly when cost per task matters more than any single benchmark score.

The key figures are twofold: DeepSeek reports a score of 77.0 on DeepSWE v1.1 and an average cost below US$0.20 for that evaluation, while its documentation prices Search and Fetch at US$0.75 per 1,000 successful requests. The first figure is a provider-reported result and has not been independently corroborated in the sources available here; the second is an official published rate, subject to change and temporary conditions.

The announcement therefore does not prove that this checkpoint is the best choice for every agent or that the savings will apply across all workloads. It does provide specifications, MIT-licensed weights, and concrete prices to test. A sensible decision is to evaluate the version against your own workload, account for tokens and search calls separately, and review the current terms before putting it into production.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.