DeepSeek V4.1 Flash 1017: benchmarks and pricing

§1The release, verified

DeepSeek announced DeepSeek-V4.1-Flash-1017 on 17 October 2026. Its changelog says the website, app and API have moved to the new version. API users should request deepseek-flash; the calling method and pricing are unchanged. The older deepseek-v4-flash and deepseek-v4-flash-vision-exp names now route to this model too. Release announcement · API model mapping.

This is a source review of the release. I have not reproduced these benchmarks or run an application migration. The results below are DeepSeek’s reported measurements; the evaluation plan at the end describes what I would test next.

My interest is the combination of better task completion and less work per task. If the same agent reaches an accepted result with fewer generated tokens and less elapsed time, the improvement can matter more than a small change in a leaderboard score.

§2Where the reported gains are strongest

The model card compares 1017 with the preceding V4.1 Flash checkpoint. The table preserves the benchmark versions. Changes are percentage points, calculated from DeepSeek’s figures. Official model card.

DeepSWE reaches 77.0 Pass@1, using the Mini-SWE-V2 harness. Terminal-Bench 4.0 uses Claude Code. Those harness details belong beside the scores: tool access, orchestration and retry policy help determine what a model can accomplish. DeepSeek’s evaluation announcement · Evaluation notes.

The release also reports HLE without tools at 42.6, with 44.0 on its text-only subset; Chartography at 83.0 and BabyVision at 92.5 with tools; and ZeroBench-main at 51.0 with tools. The changelog specifies Pass@5 for ZeroBench-main, so that result should not be read as single-attempt accuracy.

§3A strong agent result, with visible limits

DeepSeek’s own comparison shows a mixed picture. These are selected rows from the same vendor-published table, not an independent ranking. The original labels “K3” and “Opus 5” are retained to avoid implying more precise model versions than the source provides. Full comparison and footnotes.

Flash 1017 leads these competitors on DeepSWE and Automation-Bench. Opus 5 leads on Terminal-Bench, ProgramBench and NL2Repo-Bench. The practical conclusion is that 1017 deserves a workload-specific trial. This table does not establish that it is the best model for every coding or agent task.

§4What the Pareto claim actually means

A Pareto improvement means improving at least one measured objective without worsening the others in the comparison. Here, DeepSeek claims higher benchmark performance alongside shorter task runtimes and shorter outputs, and publishes a DeepSWE score-versus-cost chart. Performance–cost announcement.

The textual release notes do not provide one universal percentage reduction for output length or task time. I would not turn the chart or the headline into a blanket speed multiplier. Shorter outputs can reduce token spending, but total job time also includes tool calls, retries, queueing and verification.

The metric I would track is cost per accepted task: all model and tool charges across successful and failed attempts, divided by the number of tasks that meet the acceptance criteria. A cheaper attempt that needs repeated repairs may be a more expensive workflow.

§5Model pricing and the separate search bill

Prices below are DeepSeek’s direct API list prices checked on 17 October 2026, in US dollars per million tokens. They are not third-party hosting prices. Models and pricing.

Peak windows are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. In Dubai that is 05:00–08:00 and 10:00–14:00. Other hours, weekends and those holidays receive off-peak pricing. October 17 is a Saturday.

For a simple illustration, 100,000 uncached input tokens plus 20,000 output tokens cost $0.027 off-peak or $0.054 at peak. This arithmetic excludes search, fetching, retries and other infrastructure; it is not a measured agent-task bill.

Search and Fetch each cost $0.75 per 1,000 successful requests. Search returns structured web results; Fetch returns a page as Markdown. A successful Search must return a non-empty results array, and a successful Fetch must return content. Failed requests, empty searches and timeouts are not billed. Asking Search for more results does not change its per-request price. Search & Fetch billing rules.

One search followed by three fetches is four billable requests: $0.003. At that pattern, 1,000 research tasks would incur $3 in tool charges, before model tokens.

There is a transition rule for server-side tool calls through DeepSeek Harness, Claude Code and Codex when using the DeepSeek API. Those calls are currently free; billing starts 31 October 2026 at 16:00 UTC, or 20:00 in Dubai (1 November at midnight in Beijing). This is not a general promise that every direct Search or Fetch request is free until November. Billing start and scope.

§6Project-building improvements and open weights

DeepSeek attributes better end-to-end project development to reinforcement learning with self-feedback. The announcement demonstrates 3D scenes and games, including a Forbidden City model, a browser racing game and a Godot recreation inspired by Portal 2. These are selected vendor demonstrations, not evidence that arbitrary project requests will succeed unattended. Project-building demonstrations.

The release names optimizations for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and recommends DeepSeek Harness 0.2.1-rc.1 or later. When evaluating this claim, record the harness version as well as the model: an orchestration change can affect the result.

The official Hugging Face repository provides the weights under MIT licensing. The release also links to ModelScope. I verified the Hugging Face model card; ModelScope availability here is attributed to DeepSeek’s announcement.

Local deployment still needs a supported runtime and a memory budget. The card points to reference inference code and a Python prompt encoder, and explicitly says this release does not include a Jinja-format chat template. It also links to the maintained deepseek-recipe encoding toolkit. Open weights alone do not establish that a given desktop setup can run the model. Start with the memory-estimation method and the checkpoint’s actual files.

§7How I would evaluate it in an agent workflow

1. Freeze a small task set. Include a repository bug fix, a terminal task, a new project, a sourced research question and a tool-failure recovery task. Define acceptance checks before running either model.
2. Record the exact configuration. Capture provider, returned model identity, harness version, reasoning effort, context, tool permissions and spending limits. The deepseek-flash alias now points at 1017; an older alias alone cannot reproduce the predecessor.
3. Run repeated attempts. Use clean workspaces and equal budgets. Keep failed attempts in the record. A single polished demo cannot establish a success rate.
4. Measure the full job. Log elapsed time, input and output tokens, cache hits, billable tool calls, retries, test results and human repair time.
5. Review accepted work. For code, inspect the diff and execute the acceptance tests. For research, check that citations support the claims. Compare cost per accepted result and the time a human spends repairing it.

That is the decision the release makes interesting: whether the stronger checkpoint produces more accepted work for the same budget in the harness you actually use. The SGLang and vLLM comparison explains why configuration belongs in any performance claim.

§8Sources and evidence boundary

Researched on 17 October 2026 from DeepSeek’s official announcement, changelog, API documentation and model card. Benchmark scores and training claims are attributed to DeepSeek. Price examples and score differences are calculations from the cited inputs. No independent benchmark reproduction or live API trial was performed for this article.

• Release announcement: DeepSeek-V4.1-Flash-1017
• October 17 changelog and evaluation footnotes
• Hugging Face model card, benchmark comparison and license
• Direct API model mapping and pricing
• Search & Fetch guide, billing and limits