DeepSeek Flash 1017 brings coding gains and a new search API
DeepSeek updated V4.1 Flash on October 17 and opened its own web search service to developers. The Flash 1017 release reaches the official app, website and existing Flash API route. DeepSeek’s release record also recommends Harness version 0.2.1 rc.1 or later for the updated model.
The practical questions are whether the revised model completes useful work more reliably, how much an entire task costs and how search calls change that calculation. The published documentation provides concrete answers about billing and integration, alongside performance claims that still need independent testing.
What the reported gains actually measure
DeepSeek reports 77.0 on DeepSWE v1.1, compared with 74.2 for September’s release, and 42.3 on Terminal Bench 4.0, up from 31.2. The company identifies Mini SWE V2 for DeepSWE and Claude Code for Terminal Bench.
The company also says task runtime and output length have fallen. That is a claim about complete workloads, rather than an announced reduction in the token price. Evaluate consistent tasks and permissions. Shorter answers help only when the result remains correct.
Read both price columns before estimating costs
The pricing page lists Flash input without a cache hit at $0.15 per million tokens during off peak hours and $0.30 during peak hours. Output costs $0.60 or $1.20 respectively. Cached input costs $0.003 or $0.006. Peak periods run from 01.00 to 04.00 UTC and 06.00 to 10.00 UTC on weekdays, excluding Chinese public holidays. Weekends are entirely off peak.
V4 Pro remains separately listed as its August 13 version. The compatibility routing applies to the retired Flash and Flash Vision names, so an existing Pro integration should not be described as automatically receiving this Flash update.
Search introduces a separate usage meter
Search and Fetch each cost $0.75 per 1,000 successful requests. For existing server side tool calls through DeepSeek Harness, Claude Code and Codex, the pricing page specifies billing from November 1 at midnight Beijing time, equivalent to October 31 at 16.00 UTC.
The Search reference accepts queries up to 128 characters and lets callers request between one and 50 results. Each result includes a URL and title, a short snippet and a longer summary intended for a model. Page content can also be requested, but the content field can be empty when unavailable. Publication dates may be unknown.
That structure lets an application select a compact result list before passing larger source material to a model. It does not make every returned statement trustworthy. Developers should retain the original URLs, check whether a page supports the answer and handle missing dates or content explicitly. A relevance score is a ranking signal rather than independent verification.
The usage guide describes separate account limits of five queries per second for Search and Fetch, with further domain limits for Fetch. Successful billing requires a response with nonempty results or content. Failures, empty searches and timeouts are not charged. The guide still describes the integration billing date as forthcoming, while the pricing page already gives November 1. Teams should check the current pricing notice before deployment.
Two integration details can change the outcome
DeepSeek’s thinking mode documentation defaults to high effort. Low and maximum effort are also available. Some familiar sampling controls have no effect in thinking mode, so copying another provider’s configuration does not guarantee equivalent behavior.
For tool enabled conversations in thinking mode, the documentation requires earlier reasoning content to be returned in subsequent requests, including turns without a tool call. Omitting it can produce a 400 error. An application that only preserves the final answer can therefore fail even when its model name and credentials are correct.
The context cache guide says caching is automatic but operates on persisted matching prefixes. A shared opening does not always produce an immediate hit. Construction takes seconds, hit rates are not guaranteed and unused entries are normally cleared within hours or days. Measure the returned cache hit and cache miss token counts rather than assuming the cheapest input rate applies to every request.
What to check before moving a coding workflow
Run a representative set of tasks and record successful completion, failed tool calls, elapsed time, input and output tokens, cache hits and search requests. Keep the model revision and agent framework in the record. This separates a model improvement from a configuration change and makes the cost of retries visible. These are evaluation recommendations, not results from ByteForward testing.
The update also sits alongside DeepSeek’s broader infrastructure work, including its published Ascend kernels and TileLang support. That earlier software release concerns inference infrastructure. Today’s API checkpoint and search service concern what application developers can access and pay for. They should be assessed on their own documented scope.
Illustrative coding photograph by TREEDEO.ST on Pexels, used under the Pexels license. The image does not show DeepSeek software.
