DeepSeek-V4.1-Flash-1017: Prices, Benchmarks and the Search and Fetch Billing Date to Watch

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17. It is live in the API, the weights are on Hugging Face under the MIT license, and DeepSeek opened its Search and Fetch APIs to everyone with a billing start date that matters if your agent already calls them.

Everything below comes from DeepSeek's release note, the model card and the pricing page. Benchmark numbers are DeepSeek's own. I did not run the model.

What DeepSeek says changed

The model card calls it a minor update to DeepSeek-V4.1-Flash with improvements in reasoning, coding, agentic tasks and visual benchmarks. DeepSeek says it also reduces task execution time and output length, which it frames as a better performance and cost trade-off across agent harnesses. The release note says it is optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and recommends DeepSeek Harness v0.2.1-rc.1 or later.

The release note also describes reinforcement learning with self-feedback to improve end-to-end project development, and shows demos such as a racing game and a 3D scene. Those are vendor demos, not measurements.

Benchmarks, as DeepSeek reports them

From the model card, DS-V4.1-Flash-1017 versus the previous Flash and the models DeepSeek chose to compare:

Benchmark	1017	Previous Flash	GPT-5.6 Sol	Opus 5
Terminal-Bench 4.0	42.3	31.2	39.9	51.8
DeepSWE v1.1	77.0	74.2	73.0	74.0
ProgramBench	29.1	20.3	23.0	37.0
NL2Repo-Bench	68.2	65.4	56.8	75.3
HLE without tools	42.6	36.8	44.5	56.3

The card also lists Kimi K3 and GLM-5.3 columns, and several vision benchmarks. Two footnotes matter: Terminal-Bench 4.0 uses the Claude Code harness and DeepSWE uses the Mini-SWE-V2 harness, and the HLE figures in brackets are for the text-only subset.

Read it fairly: the new Flash beats the old Flash on every row shown, and beats the listed GPT-5.6 Sol on some rows. On the rows above, Opus 5 is ahead on all but DeepSWE. These are DeepSeek's choices of benchmarks and harnesses.

Pricing

The pricing page lists deepseek-flash (model version DeepSeek-V4.1-Flash-1017) next to deepseek-v4-pro (DeepSeek-V4-Pro-0813). Per 1M tokens, Flash:

• Input, cache hit: $0.003 off-peak, $0.006 peak.
• Input, cache miss: $0.15 off-peak, $0.30 peak.
• Output: $0.60 off-peak, $1.20 peak.

Off-peak rates are half of peak. Per DeepSeek, peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Context length is 1M, max output 384K, concurrency limit 2,500 for Flash. Flash supports vision, and the pricing page lists Pro as not supporting it.

Model names matter here. The page says use deepseek-flash, and the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but those models are retired: requests are served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price. If you pinned an old name, your traffic may already be on the new model.

Search and Fetch APIs

DeepSeek says its Search and Fetch APIs are now open to everyone, running on its in-house search engine. Each endpoint is $0.75 per 1,000 successful requests, per the pricing page.

The date to note: calls made as server-side tools in DeepSeek Harness, Claude Code and Codex are currently free for DeepSeek API users. The release note says billing starts October 31, 2026 at 16:00 UTC. The pricing page says 00:00 Beijing time on November 1, 2026, which is the same instant. If your agent does a lot of web lookups, estimate the cost before then.

Weights and license

The model card says the repository and weights are MIT licensed. It lists 1M context and recommends temperature 1.0, top_p 1.0 and max_tokens of at least 256K. There is no Jinja chat template. The release includes a Python reference encoder and a Rust-based toolkit called deepseek-recipe for prompt encoding and output parsing. If you self-host, plan time for the prompt format.

What I would check

These are my own untested suggestions.

1. Run your own task set on deepseek-flash against your current model, with your harness, and compare cost per completed task, not per token.

2. Check which model name your code sends.

3. Add up Search and Fetch calls per task and multiply by $0.75 per 1,000 for each endpoint.

4. Check whether peak hours fall in your traffic window.

What I would not conclude

• That it beats the closed models overall. DeepSeek's table shows the rows where it does and does not.
• That benchmark gains transfer to your codebase.
• That off-peak pricing applies to your workload. It depends on your hours.

Frequently asked questions

What is DeepSeek-V4.1-Flash-1017?

An update to DeepSeek-V4.1-Flash, live in the DeepSeek API as deepseek-flash, with open weights.

What license are the weights under?

MIT, per the model card.

How much does it cost?

For Flash, off-peak $0.003 per 1M cached input tokens, $0.15 uncached input and $0.60 output. Peak rates are double.

When are peak hours?

01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays.

Are Search and Fetch free?

Calls through DeepSeek Harness, Claude Code and Codex are free for now. Billing starts October 31, 2026 at 16:00 UTC, at $0.75 per 1,000 successful requests per endpoint.

Do old model names still work?

Per the pricing page, the legacy Flash names are accepted and served by the new model at Flash price.

Can I self-host it?

The weights are public, with a recommended 1M context. There is no Jinja chat template, so use the provided encoding or deepseek-recipe.

Sources

• DeepSeek API Docs, DeepSeek-V4.1-Flash-1017 release note (October 17, 2026): features, Search and Fetch pricing and billing date.
• Hugging Face, DeepSeek-V4.1-Flash-1017 model card: benchmarks, license, sampling and encoding notes.
• DeepSeek API Docs, Models and Pricing: prices, peak hours, model names, concurrency.
• TLDR AI, October 19 issue: discovery pointer only.