DeepSeek has replaced V4.1 Flash with V4.1-Flash-1017 on the web app, mobile app, and API — existing calls now hit the new model. The model itself is unchanged under the hood: it keeps the same parameters and the same MoE architecture as before. DeepSeek's release charts have it winning the head-to-head against GPT 5.6 Sol on most of 11 benchmarks, including 77.0% on DeepSWE v1.1 at roughly ¥1 per task — the cheapest point on the cost-performance frontier, at Flash-level speed. Two new API endpoints, Search and Fetch, cost $0.75 per 1,000 successful requests each (≈¥5); inside DeepSeek Harness, Claude Code, and Codex they are free until 00:00 Beijing time on November 1, 2026. Weights are open on HuggingFace and ModelScope, and DeepSeek Harness users should upgrade to 0.2.1-rc.1 or later.
V4.1 Flash moved forward to V4.1-Flash-1017, and the switch is already live everywhere at once: the official web app, the mobile app, and the API all serve the new version. It is a version bump of the Flash line, not a new generation — though commenters joked that with jumps like this, it deserves the V4.2 name.
The rollout also clears up the week's confusion: the downtime two days ago came from the Flash update, not from V4.1 Pro, as the source report notes. The API side needs no migration — per the official change log, the calling method and pricing remain unchanged, and the deepseek-flash alias now serves the new model.
Three claims from the release stand out:
- An open-weights Flash-tier model whose own release table trades blows with GPT 5.6 Sol.
- Speed as the headline — with GPT output widely reported crawling at around 20 tokens per second, a near-Sol model at Flash throughput becomes the obvious daily driver for long agent tasks.
- A self-built Search API is now open to everyone, with Fetch alongside it.
How close is V4.1-Flash-1017 to GPT 5.6 Sol, really?
On DeepSeek's own release table, V4.1-Flash-1017 wins the head-to-head against GPT 5.6 Sol on most of the 11 benchmarks and posts the best score of any listed model on 4 of them — Claude Opus 5 still tops 6. The official change log calls out three headline rows with exact scores: HLE at 42.6, Terminal-Bench 4.0 at 42.3, and DeepSWE v1.1 at 77.0. The table's footnotes name the evaluation settings: the Claude Code framework for Terminal-Bench 4.0, Mini-SWE-V2 for DeepSWE, and Pass@5 for ZeroBench-main.
The cost side is the louder claim. On the DeepSWE v1.1 cost-performance chart below (mini-swe-v2 framework, leaderboard dated 2026-09-22), the new model's max setting reaches 77.0% Pass@1 at roughly ¥1 per task — the cheapest point anywhere on the frontier, where competitor frontier points span roughly ¥2 to ¥100+ (USD converted at 6.6862 CNY, 2026-10-16). DeepSeek also says it shortened task runtimes and output lengths, so the same job finishes faster and costs less, calling the result a new performance-cost Pareto optimum across multiple harnesses.
Read these as launch claims: they are provider-reported charts, not independently reproduced results. The evaluation settings above are the ones DeepSeek disclosed per row.
How much does the Search & Fetch API cost — and when does billing start?
Both endpoints are listed at $0.75 per 1,000 successful requests (≈¥5), billed per successful request rather than by tokens or returned results — failed requests, empty search results, and timeouts are not charged. Inside DeepSeek Harness, Claude Code, and Codex, Search and Fetch reached through server-side tool calls are free for a two-week window — per-request billing starts at 00:00 Beijing time on November 1, 2026.
| Endpoint | Direct API price | Via server-side tools (Harness / Claude Code / Codex) | What it does |
|---|---|---|---|
| Search | $0.75 / 1,000 successful requests (≈¥5) | Free until 2026-11-01 00:00 (Beijing time), then billed per request | AI-native web search: multilingual, real-time results |
| Fetch | $0.75 / 1,000 successful requests (≈¥5) | Free until 2026-11-01 00:00 (Beijing time), then billed per request | Page retrieval for agents and pipelines |
Prices as stated in the official Search & Fetch guide (checked 2026-10-17).
The search layer behind them is DeepSeek's own engine, which the company says has powered search on both its web and API products all along — hundreds of millions of queries a day, an index of hundreds of billions of pages, with filtering aimed at SEO spam, GEO manipulation, and low-quality AI-generated content.
Want to try the latest DeepSeek V4.1 Flash 1017 model for yourself? Head to our coding plan and put it to work in your everyday coding workflow.
Do you need to change anything?
SaveMyToken recommendations:
- Web and mobile app users — nothing to do. Both surfaces already serve V4.1-Flash-1017.
- DeepSeek Harness users — upgrade to 0.2.1-rc.1 or later. The official recommendation; the release is tuned for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.
- API users who pin exact model IDs — update the ID. The model is published as DeepSeek-V4.1-Flash-1017; if you call the deepseek-flash alias instead, the official docs confirm it already serves the new version.
- Claude Code / Codex on a DeepSeek API key — audit Search/Fetch usage now. Server-side tool calls are free only until November 1, 2026, then bill at $0.75 per 1,000 successful requests.
- Cost watchers — measure, don't assume. The release claims shorter outputs and faster runs; check your own task costs before booking the savings.
Where are the open weights?
The model is published under DeepSeek-V4.1-Flash-1017 on both major hubs: