DeepSeek V4.1 Flash 1017: benchmarks, pricing and what changed

Sources checked: October 20, 2026.

DeepSeek V4.1 Flash 1017 is the October 17, 2026 update to DeepSeek V4.1 Flash. DeepSeek says the new checkpoint is live on its website, app and official API, with stronger project-development results and shorter outputs. The official API keeps the deepseek-flash name. The benchmark figures below are DeepSeek's published results; they are not measurements made by SeedRouter. Official announcement.

For an existing user, the main questions are what changed from the September checkpoint, whether the API configuration changes, and what the update means for task cost. Those questions can be answered without treating a release announcement as a guarantee about every workload.

What changed in DeepSeek V4.1 Flash 1017?

The release focuses on end-to-end project development, including coding and 3D-scene examples. DeepSeek attributes the improvements to reinforcement learning with self-feedback. Its announcement also emphasizes shorter output and faster task completion. Those are the developer's claims; the release examples have not been reproduced for this article.

The new checkpoint has its own official Hugging Face repository. Use that model card when checking October results. A link to the original V4.1 Flash repository can still lead readers to the September numbers, even when an article is discussing 1017.

For DeepSeek Harness users, the announcement recommends v0.2.1-rc.1 or later. That recommendation concerns the client experience with the new model; it does not mean every API user must install Harness.

DeepSeek V4.1 Flash 1017 benchmark results

The dedicated model card presents the old and new checkpoints on the same named evaluations. That makes this table more useful than comparing unrelated scores taken from separate articles.

Source: DeepSeek's 1017 model card. These are developer-reported scores. Terminal-Bench uses the Claude Code harness, while DeepSWE uses Mini-SWE-V2, according to the published evaluation notes.

The Terminal-Bench difference is 11.1 score points; DeepSWE rises by 2.8 points. Those differences describe the named tests and configurations. They do not mean that every terminal task is 11.1% faster, or that a repository task will cost 2.8% less.

Be careful with benchmark versions. A score from Terminal-Bench 2.1 is not directly comparable to this Terminal-Bench 4.0 result. The task set and evaluation environment matter, so keep the benchmark version and harness attached to a score when quoting it.

How to use the benchmark table for your own decision

The published improvements are a reason to reevaluate representative tasks. They do not establish that the new checkpoint is best for every application.

For a coding workflow, keep the repository, instructions, tool permissions and acceptance checks stable. Compare whether the task is completed correctly before comparing speed or token use. If the task requires follow-up prompts, include those turns in the cost. Otherwise, a short first answer can appear economical while moving work into later requests.

Also record the model identity the service documents. A stable API alias can point to a newer checkpoint over time. If you have no saved run from the September version and cannot access that version again, describe your result as a current-model evaluation rather than a controlled before-and-after test.

Did DeepSeek V4.1 Flash 1017 lower the API price?

DeepSeek's October 17 changelog says the API calling method and pricing remain unchanged. The announcement's lower-task-cost claim is about efficiency, including shorter output, rather than a newly announced reduction in the token unit price.

These two measurements answer different questions. A token price tells you the charge for a given quantity of input or output. Task spend depends on how many tokens and turns the work consumes. A model may finish a task with fewer output tokens while keeping the same per-token rate; a difficult task may still need more work.

Use DeepSeek's current pricing page for its official rates and billing conditions. For requests through SeedRouter, use the DeepSeek pricing guide and the service's current rates. The official release does not establish which checkpoint a separate service is serving.

Which API model name should you use?

DeepSeek's official documentation uses deepseek-flash for the current Flash model. The pricing page identifies its served version as DeepSeek-V4.1-Flash-1017. It also says the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp remain accepted and are served by the new checkpoint.

An existing direct DeepSeek integration using deepseek-flash therefore does not need to invent a new model ID containing 1017. Keep the documented request format and review behavior on your own tasks.

Other services can expose their own documented model names. The existing SeedRouter Claude Code and Codex setup guide records its configuration and September 28 tests. That article is not evidence that SeedRouter's serving checkpoint has been verified as 1017. This release review makes no such claim.

Where are the 1017 weights?

The official repository is deepseek-ai/DeepSeek-V4.1-Flash-1017. Its model card identifies the release and MIT license. Follow that repository's loading and deployment material rather than treating the availability of weights as proof that they fit your hardware.

The model release coincided with the opening of standalone Search and Fetch APIs. Those are a separate integration and billing topic, covered in the DeepSeek Search API guide. Downloading model weights does not by itself provide access to DeepSeek's hosted web-search service.

Frequently asked questions

Is 1017 a new model family?

DeepSeek presents it as an update to V4.1 Flash. The dedicated checkpoint name distinguishes the October 17 release from the earlier V4.1 Flash version.

Do I need to change deepseek-flash in the official API?

No change is described for that alias. The current official documentation maps it to 1017. Check the documentation of any other service separately.

Does the benchmark improvement guarantee a cheaper coding session?

No. A session's cost depends on input, output, cache behavior and the work needed to finish. The published scores and efficiency claims are useful context, but they do not replace a measurement on your task.

Evaluate the checkpoint against a fixed task

DeepSeek V4.1 Flash 1017 has documented improvements over the September checkpoint on several named coding evaluations. Keep those results attached to their test versions and harnesses, then evaluate the current model against a task with clear acceptance checks. Separate token pricing, total task spend and model identity when deciding whether the update changes your workflow.