DeepSeek-V4.1-Flash API

DeepSeek-V4.1-Flash-1017 at a glance

Release date

2026-10-17, on DeepSeek’s website, app and API the same day; it replaces the 2026-09-10 V4.1-Flash under the same deepseek-flash model ID, with the calling method and price unchanged

What changed

DeepSeek calls it a minor update with higher benchmark scores, faster task completion and shorter outputs, which it describes as a Pareto improvement in performance and cost; it is tuned for DeepSeek Harness (v0.2.1-rc.1 or later recommended), Tencent CodeBuddy and WorkBuddy

Model and weights

A multimodal MoE on DeepSeek’s Causal Encoder-Decoder (CED) architecture; the V4.1-Flash technical report gives 552B backbone parameters with 16B active per token in decoding, and the Hugging Face weight files total 763B parameters; code and weights are MIT-licensed

Context and output

1M-token context and up to 384K (393,216) output tokens per request; text and image input, text output; the model card recommends max_tokens of at least 256K

Thinking and effort

Thinking is on by default at high effort; reasoning_effort accepts low / high / max; thinking: {type: "disabled"} turns it off, and on the Responses API reasoning.effort "none" does the same

Official price

$0.30 input / $1.20 output per million tokens and $0.006 for cache hits at peak, half that off-peak ($0.15 / $0.60 / $0.003). Peak is 01:00–04:00 and 06:00–10:00 UTC on weekdays

API and tools

DeepSeek offers OpenAI Chat Completions, Responses and Anthropic Messages formats with tool calling, JSON output and FIM completion (non-thinking only); Wokey serves /v1/chat/completions, /v1/responses and /v1/messages

Checked 2026-10-18 · Sources: DeepSeek · Change log · Pricing · Hugging Face · arXiv · Artificial Analysis · DeepInfer

Official benchmarks: V4.1-Flash-1017 vs V4.1-Flash vs Kimi K3 vs GLM-5.3 vs GPT-5.6 Sol vs Opus 5

From DeepSeek’s model card on Hugging Face, matching its 2026-10-17 change log; these are vendor-reported results, and “—” means no score was published. Opus 5 here is the earlier Claude Opus 5, not Opus 5.5 (2026-09-22), and GPT-5.6 Sol is not GPT-6.1 Sol.

The HLE row is without tools; on the text-only subset DeepSeek also reports 44.0% for 1017 and 39.1% for the previous version, and its GLM-5.3 figure (42.0%) is text-only too, so that cell is blank. Terminal-Bench 4.0 ran in the Claude Code harness, DeepSWE v1.1 in Mini-SWE-V2, and ZeroBench at Pass@5. Different vendors’ tables do not compare cell by cell: Moonshot’s own card gives Kimi K3 77.8% on ProgramBench against 17.5% here, and DeepSeek gives GLM-5.3 48.8% on AutomationBench v1.0.6 against 48.2% on Z.ai’s page. The DeepSWE v1.1 figures for K3 (67.5%) and GLM-5.3 (66.9%) match Z.ai’s table.

Which to use: V4.1-Flash-1017, Claude Opus, Kimi K3, GLM-5.3, same-tier models or V4 Pro

Upgrading from V4.1-Flash (0910)

Nothing to change: deepseek-flash switched to 1017 on 2026-10-17 at the same price. In DeepSeek’s table the biggest gains are on agentic and terminal tasks: Terminal-Bench 4.0 from 31.2% to 42.3%, ProgramBench from 20.3% to 29.1%, Apex Bench v1.1 from 64.6% to 72.1%, DeepSWE v1.1 from 74.2% to 77.0%. DeepSeek also says it finishes tasks faster with shorter outputs, and reasoning bills as output, so a task costs less; no independent test has confirmed the shorter outputs yet.

vs Claude Opus 5 / Opus 5.5

Even in DeepSeek’s own table Opus 5 leads the hardest rows: HLE 56.3% vs 42.6%, Terminal-Bench 4.0 51.8% vs 42.3%, ProgramBench 37.0% vs 29.1%, NL2Repo-Bench 75.3% vs 68.2%, plus all three vision benchmarks; 1017 leads on DeepSWE v1.1 (77.0% vs 74.0%), AutomationBench (60.5% vs 50.3%), Agents’ Last Exam and Apex Bench. Opus 5.5, released 2026-09-22, is not in the table. Opus 5.5 lists at $4 / $20, about 13× 1017’s peak input price and 17× its peak output price. Pick Opus for the hardest builds and work that has to be right the first time; use 1017 for high-volume agent loops, sub-agents and routine coding where cost per task matters.

vs Kimi K3 and GLM-5.3

DeepSeek’s table puts 1017 ahead of both on every coding and agent row, for example DeepSWE v1.1 77.0% vs 67.5% (K3) and 66.9% (GLM-5.3), and Terminal-Bench 4.0 42.3% vs 12.6% and 37.9%; only on HLE is K3 slightly higher (43.5% vs 42.6%). The price gap is larger: K3 lists at $3 / $15 and GLM-5.3 at $1.4 / $4.4, against $0.30 / $1.20 for 1017 at peak. GLM-5.3 is text-only and K3 also takes video; both 1017 and K3 take images. These are DeepSeek’s numbers, so check Z.ai’s and Moonshot’s own tables before deciding.

vs GPT-6 Luna, GLM-5.3-Flash and Claude Haiku 5.5

For high-volume work the real alternatives are these same-tier models, and all of them are on Wokey. Artificial Analysis Intelligence Index v4.3.2 has only the September V4.1-Flash so far: 39 at max effort for $0.27 per task; on the same version Claude Haiku 5.5 scores 43 ($0.21 per task), GLM-5.3-Flash 42 ($0.28) and GPT-6 Luna 38 ($0.07, the cheapest per task). AA has not scored 1017 yet. Official list prices: Luna $0.10 / $0.50, GLM-5.3-Flash $0.15 / $0.50, Haiku 5.5 $0.10 / $0.50 for prompts up to 100K tokens and five times that above; DeepSeek $0.30 / $1.20 at peak and $0.15 / $0.60 off-peak. DeepSeek has the cheapest cache hits ($0.003–$0.006, against $0.01 for Luna and Haiku 5.5 and $0.03 for GLM-5.3-Flash), which is the largest cost in agent loops that resend a long prefix. AA measured output speed at about 235 tokens/s for Haiku 5.5, 176 for DeepSeek, 120 for Luna and 71 for GLM-5.3-Flash. Pick Luna for the lowest cost per task, Haiku 5.5 or GLM-5.3-Flash for the top score in this tier (Haiku 5.5 long prompts cost five times as much), and 1017 for long, cache-heavy agent loops or when you want open weights.

vs DeepSeek V4 Pro

DeepSeek has not released a V4.1 Pro. deepseek-v4-pro is still the 0813 snapshot at $1.32 / $3.96 at peak, 4.4× Flash on input and 3.3× on output, without image input, and with a concurrency limit of 500 against Flash’s 2,500. DeepSeek publishes no Pro scores beside 1017; Z.ai’s GLM-5.3 table lists V4 Pro 0813 at 62.7% on DeepSWE v1.1, below the 77.0% DeepSeek reports for 1017, though the harnesses may differ. Use deepseek-flash for new work.

Choosing effort

Thinking defaults to high. DeepSeek maps requested levels onto three: minimal and low → low; medium, high and xhigh → high; max and ultra → max. Keep the default high for coding and agents, use max for the hardest reasoning and low for extraction and short answers, and turn thinking off (thinking: {type: "disabled"}) for FIM, classification or latency-bound calls. Reasoning tokens bill at the output rate.

DeepSeek thinking-mode docs

Third-party reviews

• Kingy AI: A small pilot on 2026-10-17 through Vercel AI Gateway: seven fixed tasks with deterministic graders, each run twice, against GPT-6.1 Sol, Claude Sonnet 5.5 and Qwen3.8-27B. 1017, Sol and Sonnet all passed 14 of 14, so the battery could not rank them. 1017’s clearest advantage was cost: about $0.016 against roughly $0.058 for Sol and $0.065 for Sonnet. Speed was mixed: a 2.47 s median on ordinary tasks (Sonnet 1.59 s, Sol 4.77 s), but 27–28 s on one interval-subtraction task that Sonnet finished in about 3 s. The author did not verify the shorter-output claim or test agent harnesses.
• Paweł Huryn (Bug Hunt Bench): Two real repositories with 105 planted bugs that frontier models missed early this year; the model has to find and fix them. 1017 at max effort scored 34 on the first run and 38.7 averaged over three, against 21.7 for the September version; in the same test Opus 5.5 (max) scores 41.7 and Gemini 4 Argon (high) 40.7. At high effort it averaged 32 over two runs, level with Opus 5.5 at high (31.7). The author notes that 1017 varies far more from run to run than other models.
• ML_Bear (X): A largely negative first take: extremely cheap (about 1/16 of Opus), fast on short outputs but slow on long ones; the fatal weakness is missing detailed instructions, with accuracy far behind Sol and Opus; weak Japanese; and no strict JSON Schema mode. The suggestion is to use it for high-volume work where accuracy matters less.
• Startup Fortune: Argues the release matters less for its scores than for its packaging: like the September V4.1-Flash launch, it shipped together with a new DeepSeek Harness version and named partner tools, steering developers toward a stack DeepSeek controls end to end. It also notes that the MIT license lets you run the model in any tool.

Using 1017 in Claude Code, and request rules to check

These rules come from DeepSeek’s API docs and model card, plus how Wokey routes deepseek-flash. Check them before switching or wiring it up for the first time.

• Keep the model ID deepseek-flash: DeepSeek’s API has no separate 1017 ID, and the previous version cannot be pinned.
• On Wokey, deepseek-v4-flash is a separate model whose first route is Volcengine’s DeepSeek V4 Flash deployment, and it does not accept images. To get 1017, switch the ID to deepseek-flash. (On DeepSeek’s own API the old name now runs 1017.)
• With tools in thinking mode, pass back reasoning_content for every previous assistant turn, including turns without a tool call, or DeepSeek returns 400 (The reasoning_content in the thinking mode must be passed back to the API). Appending the returned message object as-is is enough; without tools the field is ignored.
• In thinking mode temperature, presence_penalty and frequency_penalty are accepted but have no effect, and top_p below 0.95 is raised to 0.95; the model card recommends temperature 1.0 and top_p 1.0.
• Structured output: response_format supports only json_object (valid JSON; the prompt must mention “json” and show an example format), not json_schema. DeepSeek’s strict function calling is a beta on its /beta base URL, and Wokey calls the standard endpoint, so validate tool arguments yourself.
• Up to 384K (393,216) output tokens per request; the model card recommends max_tokens of at least 256K, and reasoning bills as output, so leave room in output budgets.
• Peak hours: DeepSeek charges full price 01:00–04:00 and 06:00–10:00 UTC on weekdays (09:00–12:00 and 14:00–18:00 Beijing time) and half price otherwise. Wokey uses the same windows, so batch work on weekends or outside those hours costs half; unlike DeepSeek, Wokey does not treat Chinese public holidays as off-peak.
• Using it in Claude Code: set ANTHROPIC_BASE_URL to https://api.wokey.ai, ANTHROPIC_AUTH_TOKEN to your Wokey API key and ANTHROPIC_MODEL to deepseek-flash. Codex CLI uses /v1/responses and OpenCode /v1/chat/completions; see the client guides above. DeepSeek Harness, the agent DeepSeek recommends, also works with Wokey; setup is at /docs/deepseek-harness.

Tool-call follow-up with reasoning_content passed back (/v1/chat/completions)

{
  "model": "deepseek-flash",
  "reasoning_effort": "high",
  "tools": [{ "type": "function", "function": { "name": "read_file", "parameters": { "type": "object", "properties": { "path": { "type": "string" } }, "required": ["path"] } } }],
  "messages": [
    { "role": "user", "content": "Why does the build fail?" },
    { "role": "assistant", "content": null, "reasoning_content": "Read the build log first.", "tool_calls": [{ "id": "call_1", "type": "function", "function": { "name": "read_file", "arguments": "{\"path\":\"build.log\"}" } }] },
    { "role": "tool", "tool_call_id": "call_1", "content": "error TS2307: Cannot find module './config'" }
  ]
}

Catalog pricing snapshot

Prices are per 1M tokens; request-time pricing controls settlement.

Official price source: https://api-docs.deepseek.com/quick_start/pricing

Search / Fetch charges and Wokey support

How are DeepSeek Search / Fetch API calls billed?

Search returns structured web results; Fetch returns one page as Markdown. DeepSeek lists each API at $0.75 per 1,000 successful requests in its English guide, or ¥5 in its Chinese guide. These are official currency-specific rates, not Wokey prices or currency conversions. Only HTTP 200 responses with results are billed: a non-empty Search result array or page content from Fetch. Failures, empty searches and timeouts are not billed; count does not change the per-request price. Search and Fetch count separately, so a search followed by several fetches is not one billable call. Model tokens are billed separately.

Which searches start being billed on November 1, 2026?

DeepSeek says server-side Search/Fetch tool calls made through DeepSeek Harness, Claude Code or Codex using the official DeepSeek API temporarily have no tool charge. Per-request billing begins at 2026-11-01 00:00 Beijing time (UTC+8, the same offset as Singapore). This waiver does not make every direct Search/Fetch API call free or waive model token charges. The official policy is not a Wokey free-usage promise.

Does Wokey deepseek-flash support standalone Search / Fetch?

Wokey provides the model inference interfaces listed on this page, not standalone /v1/tools/search or /v1/tools/fetch endpoints. Tool calling support does not imply support for every upstream built-in web tool. Server-side search within model requests depends on the actual supply route and is not currently listed as a verified, uniformly supported Wokey capability; this page publishes no Wokey per-call Search/Fetch price. For documented standalone search and fetch support, use a DeepSeek API key with api.deepseek.com, configured and billed separately from Wokey model calls.

Checked 2026-10-19 · DeepSeek · Search & Fetch (USD) · DeepSeek · Search & Fetch (CNY) · DeepSeek · Change log

Measured on Wokey

Last 28 days (2026-09-29 to 2026-10-26), updated every Monday · Real requests from 100+ accounts

Time to first token, short prompts

2.0 s (Median · prompts under 8K tokens · p90 4.6 s)

Time to first token, agent-size prompts

3.8 s (Median · prompts of 32K–128K tokens)

Prompt cache hit rate

95% (Share of input tokens read from cache)

Most-used clients: DeepSeek Harness 42%, Magpie 21%, Claude Code 10%. Share of requests from identified clients; 28% of requests had an identified client.

How this is measured

Time to first token runs from the Wokey gateway receiving the request to the first streamed content (text, reasoning summary or refusal). It includes queueing, retries on another account and the model's thinking time, so it is longer than the upstream's own first-token latency. Only successful streamed requests count, banded by prompt size (input plus cache reads and writes).

Cache hit rate = cache-read tokens ÷ input tokens over successful requests. It depends on the client and how it is used; the cache hit-rate pages break it down by client.

Enterprise accounts and Wokey's own test accounts are excluded. A model is shown once at least 20 users used it in the last 28 days and no single user sent more than 50% of its requests, counting accounts registered through one referral chain as one user; each first-token figure needs at least 500 requests.

Cache hit rate by client

Model capabilities

Context

1,000,000 tokens

Max output

393,216 tokens

Streaming

Supported

Tools

Supported

Vision

Supported

Supported API forms

/v1/chat/completions · /v1/responses · /v1/messages

Send a request

/v1/chat/completions

curl https://api.wokey.ai/v1/chat/completions \
  -H "Authorization: Bearer $WOKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-flash","messages":[{"role":"user","content":"Hello"}]}'

/v1/responses

curl https://api.wokey.ai/v1/responses \
  -H "Authorization: Bearer $WOKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-flash","input":"Hello"}'

/v1/messages

curl https://api.wokey.ai/v1/messages \
  -H "x-api-key: $WOKEY_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-flash","max_tokens":1024,"messages":[{"role":"user","content":"Hello"}]}'

Use DeepSeek-V4.1-Flash in your tools

The gateway is https://api.wokey.ai (OpenAI-compatible clients usually take https://api.wokey.ai/v1), and the model ID is deepseek-flash. Open the guide for the client you use.

• Use DeepSeek-V4.1-Flash in Claude Code: Anthropic Messages · base URL and API key setup
• Use DeepSeek-V4.1-Flash in Codex CLI: OpenAI Responses · base URL and API key setup
• Use DeepSeek-V4.1-Flash in Claude Desktop: Anthropic Messages · base URL and API key setup
• Use DeepSeek-V4.1-Flash in opencode: OpenAI Chat Completions · base URL and API key setup
• Use DeepSeek-V4.1-Flash in OpenClaw: OpenAI Chat Completions · base URL and API key setup
• Use DeepSeek-V4.1-Flash in Hermes Agent: OpenAI Chat Completions · base URL and API key setup
• Use DeepSeek-V4.1-Flash in DeepSeek Harness: OpenAI Chat Completions · base URL and API key setup
• OpenRouter alternative: Wokey vs OpenRouter: Compare token prices, payment fees, API forms, and upstream sources.
• Verifiable AI API: which models and API forms carry response proofs: Response proofs currently cover only raw passthrough Claude Messages and GPT Responses on official routes.

Claude Code

export ANTHROPIC_BASE_URL="https://api.wokey.ai"
export ANTHROPIC_AUTH_TOKEN="$WOKEY_API_KEY"
export ANTHROPIC_MODEL="deepseek-flash"

Example production call

These usage values come from a completed DeepSeek-V4.1-Flash call that succeeded on its first attempt and was reconciled with its bill.

Total input

95569

Output

1228

Cached share of input

99.8%

Call duration

8.85 s

Chat Completions usage

Compiled from the billing record into API field examples. prompt_tokens is total input, including cached tokens.

{
  "model": "deepseek-flash",
  "usage": {
    "prompt_tokens": 95569,
    "prompt_tokens_details": {
      "cached_tokens": 95360
    },
    "completion_tokens": 1228,
    "total_tokens": 96797
  }
}

Token accounting

These are the final per-million-token rates recorded for this call.

Billed amount

$0.000787

Cost = sum of tokens × their rate ÷ 1,000,000, rounded to six decimals after summing.

Source of this call

This call used the OpenCode channel at opencode.ai with API Key authorization. Usage was metered upstream and the request returned HTTP 200.

Verify Prompt Cache

1. Keep the shared prefix of long prompts identical and reuse context according to the selected API’s cache rules.
2. Check prompt_tokens_details.cached_tokens; zero means no cache reads were recorded for this call.
3. Uncached input = prompt_tokens minus cached tokens. Price each bucket separately and add the output charge.

Frequently asked questions

Related models and documentation

• DeepSeek V4 Flash
• DeepSeek V4 Pro
• API documentation
• All models and pricing
• Provider Node