DeepSeek-V4.1-Flash-1017 is out: what changed in deepseek-flash

On 17 October 2026 DeepSeek changed the model behind its deepseek-flash API name. It's now DeepSeek-V4.1-Flash-1017. The name, the way you call it and the price per token stayed the same. Only what answers your requests is new.

DeepSeek says it scores higher on benchmarks, finishes tasks faster and writes shorter answers. The independent Artificial Analysis index puts it at 45, against 39 for the previous V4.1-Flash. If your code calls deepseek-flash or the older deepseek-v4-flash on DeepSeek's API, you're already on it.

What changed, and what didn't#

DeepSeek's changelog puts it in one line for 17 October: “The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model.” It gives no end date for the old names; its 10 September entry says they're routed temporarily.

Separately, DeepSeek open-sourced its inference engine, DeepInfer, under Apache-2.0. The README lists only DeepSeek-V4.1-Flash as supported, with 1017 as the default checkpoint.

The numbers: DeepSeek's table and one outside check#

DeepSeek's announcement, “Wish-Driven Programming Made Real”, calls 1017 a “Pareto improvement in performance and cost”. The price per token didn't move, so any saving on the bill has to come from shorter answers. The announcement describes the speed and length gains only in words, with no figures in its text, so there's no number to quote.

The model card calls 1017 a “minor update” and compares it with the previous V4.1-Flash:

These are DeepSeek's own results, read from the card on 20 October 2026, not an independent test. 1017 is ahead on every row, with the widest gaps on Terminal-Bench 4.0 and ProgramBench. DeepSeek names the harness for two rows (Claude Code for Terminal-Bench 4.0, Mini-SWE-V2 for DeepSWE v1.1) and doesn't state the reasoning effort for any row.

The card also lines 1017 up against K3, GLM-5.3, GPT-5.6 Sol and Opus 5, in columns DeepSeek picked. (That GLM-5.3 isn't the GLM-5.3-Flash in our catalog.) 1017 is highest on four of the 11 rows: DeepSWE v1.1 (77.0, against 74.0 for Opus 5), Automation-Bench v1.0.6, Agents' Last Exam and Apex Bench v1.1. On the other seven Opus 5 is higher, for example 51.8 against 42.3 on Terminal-Bench 4.0 and 56.3 against 42.6 on HLE without tools.

For an outside view there's Artificial Analysis. On its Intelligence Index (v4.3.2, Max reasoning mode) 1017 scores 45 and runs at 210.6 output tokens per second. The previous V4.1-Flash scores 39 in the same mode, and its page now marks it outdated and points to 1017. It's one index in one mode, and the page doesn't say when the speed was measured, so read the figures as a snapshot from 20 October 2026.

The price didn't change#

The official price of deepseek-flash per 1M tokens, as checked on 20 October 2026 on DeepSeek's pricing page, which carries no date itself:

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, except Chinese public holidays. All other time, weekends and Chinese holidays included, is off-peak, at half the price. The context window is 1M tokens and one answer can run up to 384K. For how caching and peak hours show up in a real bill, see DeepSeek API pricing in 2026.

Search and Fetch API: billing starts 1 November#

The same day DeepSeek launched a Search API and a Fetch API on its own search engine. Each costs $0.75 per 1,000 successful requests. A request is billed only if it returns HTTP 200 and a result, so empty results, errors and timeouts cost nothing. Server-side search and fetch calls made through DeepSeek Harness, Claude Code or Codex are free for now. Per-request billing starts at 00:00 Beijing time on 1 November 2026, which is 16:00 UTC on 31 October.

The guide spells out the free period only for those server-side calls. It doesn't say whether direct API calls are billed before 1 November. The free period applies to DeepSeek's own API. It doesn't extend to requests sent through CheapTokens, Claude Code ones included.

What to check in your code#

Which release answers which name today:

• Re-run your own evals. If a prompt, parser or test relied on the old model's answer length or format, shorter answers can break it. If answers got shorter, output-token spend should drop. DeepSeek's announcement gives no output-length figure in its text, so trust your own logs.
• Move off the legacy names. They have no announced end date, and DeepSeek's own advice is to use deepseek-flash.
• Don't assume a new base URL keeps the behavior. deepseek-v4-flash on DeepSeek's API and 0731 on another endpoint are different releases.

Where DeepSeek-V4-Flash 0731 fits#

DeepSeek-V4.1-Flash-1017 isn't in the CheapTokens catalog as of 20 October 2026. The DeepSeek model that is there is DeepSeek-V4-Flash, release 0731 (31 July 2026). It's an earlier release, so the results above don't apply to it. We found no direct 0731-versus-1017 comparison from DeepSeek, and the Artificial Analysis pages we checked don't list 0731.

Which one suits you depends on the job:

• Official deepseek-flash: you need 1017 itself, 1M tokens of context, answers up to 384K tokens or vision.
• DeepSeek-V4-Flash 0731 on CheapTokens: you generate long outputs or run agent loops and want a bill you can predict. One price for input and one for output at any hour; the official 1017 doubles its price in peak hours. Context is 380K, and one answer can run up to 32,768 tokens.

Here's one workload priced both ways: 10M input tokens and 1M output tokens in a day, an input-heavy mix like the one an agent produces when it re-sends its context every turn. Token amounts below are in millions, official prices are DeepSeek's as of 20 October 2026, and 80% cache hits is an assumption, not a measurement. Your own logs show your real share.

• 10 × $0.30 + 1 × $1.20 = $4.20 — official, peak, every input token a cache miss
• 10 × $0.15 + 1 × $0.60 = $2.10 — official, off-peak, every input token a cache miss
• 8 × $0.006 + 2 × $0.30 + 1 × $1.20 = $1.85 — official, peak, 80% of input tokens are cache hits
• 8 × $0.003 + 2 × $0.15 + 1 × $0.60 = $0.92 — official, off-peak, 80% of input tokens are cache hits

On our side the same workload on DeepSeek-V4-Flash 0731 costs $0.321 at today's rates.

• Official 1017 · peak · no cache hits$4.20
• Official 1017 · off-peak · no cache hits$2.10
• Official 1017 · peak · 80% cache hits$1.85
• Official 1017 · off-peak · 80% cache hits$0.92
• CheapTokens · DeepSeek-V4-Flash 0731$0.321

The chart leaves two things out. It compares different releases, so the 1017 results above don't carry over to 0731. And it stops at 80% cache hits. Push that toward 100% with short answers and the official price can come out ahead: DeepSeek bills cache hits at $0.003 per 1M off-peak, while we bill one input price per model, $0.0221 per 1M, and one output price, with no separate cache-hit rate. Take 10M input tokens that are all cache hits and 0.1M output tokens:

• 10 × $0.003 + 0.1 × $0.60 = $0.09 — official, off-peak, every input token a cache hit
• $0.231 — DeepSeek-V4-Flash 0731 on CheapTokens

These are token prices only. Balance top-ups carry a 5% fee, which isn't in our figures.

If you run agent loops or long generations and don't want the price to move with the hour, DeepSeek-V4-Flash 0731 is worth a test on your own prompts. Top up a balance with crypto or by card through a partner, point an OpenAI-style client at https://api.cheaptokens.dev/v1 (Anthropic-style clients such as Claude Code use https://api.cheaptokens.dev without /v1), and compare the bill with yours. One balance covers every model in the catalog. The Claude Code setup is in How to connect Claude Code to CheapTokens. Current rates for every model:

Related reading#

• DeepSeek API pricing in 2026 — what you actually pay
• GLM vs DeepSeek vs MiniMax: which model to choose
• MiniMax M3.1 Flash is out — benchmarks, speed and API price