DeepSeek V4.1-Flash-1017: What Changed, Pricing & Search API

The V4.1-Flash-1017 snapshot (Oct 17) and the Search & Fetch API, verified end to end: benchmark deltas, the full off-peak/peak price table, and the Nov 1 billing change at $0.75/1k that ended the free promo.

Nov 2, 2026

On October 17, 2026, DeepSeek shipped two things at once: a new V4.1-Flash-1017 snapshot of its workhorse Flash model, and the general availability of its Search & Fetch API — the in-house search engine behind DeepSeek's agentic push. If you only skimmed the announcement, you missed the detail that matters most today: the free-server-side-tool promo ended on November 1, and Search/Fetch calls now bill at $0.75 per 1,000 successful requests.

This article is the verified field guide: what actually changed in the 1017 snapshot (with the official benchmark deltas), the complete off-peak/peak price table, how the Search & Fetch API works and what it costs now, and the exact model IDs to use — including in Claude Code via the Anthropic-compatible endpoint.

Quick answer

What it is: a new snapshot of DeepSeek-V4.1-Flash (previous snapshot: September 10), trained with reinforcement learning with self-feedback for end-to-end project development. Same architecture, same API model ID deepseek-flash, same token pricing.
Benchmark deltas (official): HLE 42.6 (up from 36.8), Terminal-Bench 4.0 at 42.3, DeepSWE v1.1 at 77.0% Pass@1.
Context: 1M tokens in, up to 384K tokens out, 2,500 concurrency limit.
Price: unchanged Flash rates — $0.15–$0.30 input and $0.60–$1.20 output per million tokens depending on off-peak/peak hours.
Search & Fetch API: GA since October 17. $0.75 per 1,000 successful requests per endpoint since November 1, 2026 — the promo that made it free through Harness/Claude Code/Codex has ended.
In Claude Code: set ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic with model deepseek-flash.

What V4.1-Flash-1017 is

The name decodes simply: V4.1-Flash is the model, 1017 is the snapshot date (October 17). DeepSeek ships dated snapshots rather than version bumps here — the previous one landed September 10 — and calling the API with deepseek-flash always routes you to the newest snapshot. The 1017 snapshot keeps the architecture and adds training: specifically, reinforcement learning with self-feedback, which DeepSeek frames under the banner "Wish-Driven Programming Made Real" — the model states an outcome, executes end to end, and grades its own work.

Two framing claims from the official announcement worth taking seriously but weighing as claims: a "Pareto improvement in performance and cost" (better benchmarks, faster task completion, shorter outputs), and optimization for DeepSeek v0.2.1-rc.1+ Harness, Tencent CodeBuddy, and WorkBuddy. The snapshot is also open-sourced on Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash-1017) and ModelScope.

The official benchmark deltas, snapshot vs snapshot:

Benchmark	V4.1-Flash (09-10)	V4.1-Flash-1017
HLE	36.8	42.6
Terminal-Bench 4.0	—	42.3
DeepSWE v1.1 (Pass@1)	—	77.0%

The September snapshot's published reference points (GPQA Diamond 90.9, Codeforces 3471) were not re-tabulated against 1017 in the changelog, so treat the table above as the complete official delta picture — impressive on agentic coding, thin elsewhere by omission.

V4.1-Flash pricing: the full table

Flash pricing did not change with the snapshot — which is the point. deepseek-flash has its own rate card, well below V4-Pro's. All prices per million tokens:

Tier	Off-peak	Peak
Cache-hit input	$0.003	$0.006
Cache-miss input	$0.15	$0.30
Output	$0.60	$1.20

The peak window is 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday (excluding Chinese holidays); off-peak is everything else, at half price. For contrast, deepseek-v4-pro runs $0.66–$1.32 input and $1.98–$3.96 output — Flash is roughly 4x cheaper at the output tier. A third-party figure of "~$0.14/$0.28" input circulates in aggregator posts; it does not match the official pricing page, so don't budget off it.

Context and limits: 1M-token context, up to 384K output tokens, 2,500 concurrent requests. Legacy model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those underlying models are retired — requests route to V4.1-Flash-1017 at Flash pricing, so old configs keep working and silently get better.

The Search & Fetch API: GA, endpoints, and the Nov 1 billing change

This is the half of the October 17 announcement with an operational deadline attached. The Search & Fetch API is DeepSeek's in-house search engine (claimed index: hundreds of billions of pages, with spam and geo-filtering) exposed as two standalone JSON endpoints — not a chat model parameter:

curl -X POST https://api.deepseek.com/v1/tools/search \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "site reliability engineering salary 2026", "count": 5, "need_content": true}'

POST /v1/tools/search — parameters: query, count, need_content, ori_question, localization. Rate limit: 100 QPS.
POST /v1/tools/fetch — parameter: url. Rate limit: 25 QPS plus a per-domain cap.
Authentication is your normal DeepSeek API key.

What changed on November 1, 2026: from launch until October 31, calls made through DeepSeek Harness, Claude Code, or Codex hit these endpoints as a free server-side tool. That promo has ended — since 00:00 Beijing time on November 1, billing applies to everyone: $0.75 per 1,000 successful requests, billed separately per endpoint (Search and Fetch each count), and only on HTTP 200 responses with non-empty results. Failed calls don't bill.

Practical read: at $0.75/1k, an agent doing 50 searches per working day costs about $0.75/month in search calls — the promo ending is not a budget emergency, but it does mean unmetered free searching inside Claude Code or Codex is over. Third-party watchers note the old third-party-search workaround used to cost around $1.80/1k, so DeepSeek's own endpoint is still the cheap route.

Using V4.1-Flash-1017 in Claude Code

The Anthropic-compatible endpoint is unchanged:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "your-deepseek-key",
    "ANTHROPIC_MODEL": "deepseek-flash"
  }
}

The model mapping matters: claude-haiku and claude-sonnet alias names route to deepseek-flash, while claude-opus* routes to deepseek-v4-pro at Pro rates — if your settings still ask for Opus, you're paying 4x without noticing. Unrecognized non-Claude model names return a 400. Set deepseek-flash explicitly and you're on the 1017 snapshot with Flash billing.

If you run the DeepSeek Harness, the snapshot is optimized for Harness v0.2.1-rc.1 and newer — worth upgrading the harness before judging the new snapshot's agentic behavior.

FAQ

What is DeepSeek V4.1-Flash-1017?

A snapshot of DeepSeek-V4.1-Flash released October 17, 2026 — the same architecture as the September 10 snapshot with substantially more training, including reinforcement learning with self-feedback for end-to-end project development. The API model ID stays deepseek-flash; old deepseek-v4-flash names route to it automatically.

How much does DeepSeek V4.1-Flash cost?

$0.15–$0.30 per million input tokens and $0.60–$1.20 per million output tokens (off-peak vs peak; peak is 01:00–04:00 and 06:00–10:00 UTC on weekdays). Cache-hit input is $0.003–$0.006 per million. Pricing is unchanged from the previous Flash snapshot.

How much does the DeepSeek Search API cost?

Since November 1, 2026: $0.75 per 1,000 successful requests, billed separately for Search and Fetch. The launch promo that made the endpoints free inside Harness, Claude Code, and Codex ended October 31. Only HTTP 200 responses with non-empty results are billable.

Can Claude Code use DeepSeek's web search?

Yes — Claude Code's server-side tool calls hit the same /v1/tools/search and /v1/tools/fetch endpoints, now billed at $0.75/1k on your DeepSeek account. There is no separate toggle to enable; it works through the Anthropic-compatible endpoint and your API key.

What model ID should I use for V4.1-Flash-1017?

deepseek-flash. That ID always routes to the newest Flash snapshot. In Claude Code's Anthropic-compatible endpoint, use it as ANTHROPIC_MODEL; note that claude-opus* aliases route to the far more expensive deepseek-v4-pro.

What is the context window?

1M tokens of context with up to 384K output tokens, and a 2,500-request concurrency limit per the pricing page.