DeepSeek-V4.1-Flash pricing, plans and limits

Cost-sensitive coding agents and long-document jobs that need a 1M-token window, image input and self-hostable weights fit this model. The October snapshot lifted DeepSeek's reported DeepSWE coding score, though Claude Opus 5 still leads DeepSeek's own table on four of the harder rows.

DeepSeek-V4.1-Flash is an open-weight model that reads text and images, launched in September 2026 and refreshed on 17 October. Its 552B-parameter Mixture-of-Experts backbone runs a 1M-token context, and Artificial Analysis rates it 39 for intelligence at max reasoning effort, well above the median for similarly priced open-weight models.

Provider: DeepSeek · Family: DeepSeek V4

More about DeepSeek on HokAI

Context window: 1,048,576 tokens · Max output: 384,000

Input modalities: text, image, tool-calls · Output: text, tool-calls

About DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash arrived in early September 2026 as the smallest model in the lab's Causal Encoder-Decoder (CED) family, replacing the V4-Flash-0731 build and the experimental Vision-Exp preview. On 17 October 2026 DeepSeek shipped a minor update named DeepSeek-V4.1-Flash-1017. Per the API changelog it sits behind the same deepseek-flash model name, with calling method and prices unchanged. This page covers the architecture and launch figures, then what the 1017 snapshot changed.

The backbone is a 552-billion-parameter Mixture-of-Experts with 1 shared expert plus 384 routed experts per layer and 6 active per token. The CED design runs an encoder stage and a decoder stage, so DeepSeek's design activates 8B parameters on input and 16B on output, which keeps per-token compute far below the headline size. Input is text and images through a DeepSeek-ViT encoder trained jointly with the text path; output is text only. The context window is 1,048,576 tokens with a 384,000-token output cap, and the API offers function calling, JSON output, the Responses API and an Anthropic-format endpoint that DeepSeek documents for Claude Code.

DeepSeek reports that 1017 raised DeepSWE v1.1 from 74.2 to 77.0, Terminal-Bench 4.0 from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, NL2Repo-Bench from 65.4 to 68.2 and Humanity's Last Exam without tools from 36.8 to 42.6 (44.0 on the text-only subset). Terminal-Bench 4.0 ran inside the Claude Code scaffold and DeepSWE inside Mini-SWE-V2. The vendor also says tasks now finish with shorter runtime and fewer output tokens. The GPQA Diamond 90.9 and Terminal-Bench 2.1 90.6 figures on this page are the September launch numbers, which DeepSeek did not republish for 1017. Every score here is vendor-run and has not been independently reproduced.

Independent measurement is more mixed. Artificial Analysis rates the model (at max reasoning effort) 39 for intelligence on index v4.3.2, 231.2 output tokens per second and a $0.18 blended price per 1M tokens, but also records 250M output tokens to run the index against a median of 140M for comparable models. The per-token price is low; the model talks a lot, so cost per finished task is the number to test. Search and Fetch APIs were opened on 17 October at $0.75 per 1,000 successful requests each. For a wider price comparison see the cheapest LLM API guide, and for coding-tool context the best AI coding assistants guide.

Where the alternatives win: in the table DeepSeek published beside the 1017 weights, Claude Opus 5 scores higher on HLE without tools (56.3 against 42.6), Terminal-Bench 4.0 (51.8 against 42.3), ProgramBench (37.0 against 29.1) and NL2Repo-Bench (75.3 against 68.2), and GPT-5.6 Sol edges it on ZeroBench-main (53.0 against 51.0). Kimi K3 trails it on every row in that table, as does GLM-5.3. Teams that need audio or video input, a published system card or formal compliance paperwork should price those alternatives, and teams that want the older, larger sibling can still call DeepSeek-V4-Pro, which DeepSeek said at launch it would keep serving after mid-September 2026.

Weights are on Hugging Face under the MIT License, so commercial self-hosting and fine-tuning need no separate vendor licence. The model is also listed on OpenRouter, Azure AI Foundry, Vertex AI, AWS Bedrock and Fireworks AI. Neither the pricing page, the changelog nor the model card states a training or retention policy for API traffic, and none links a system card, so review DeepSeek's Open Platform terms before sending sensitive data.

Pricing

Off-peak (every hour outside the windows below): $0.15 per 1M input tokens on a cache miss, $0.003 on a cache hit and $0.60 per 1M output tokens. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays, and double every rate to $0.30 / $0.006 / $1.20. DeepSeek's pricing page lists no free API tier (the consumer chat app is free separately). Search and Fetch are billed separately at $0.75 per 1,000 successful requests each; server-side calls made through Claude Code or Codex integrations are free until 00:00 Beijing time on 1 November 2026. For comparison, [DeepSeek-V4-Pro](/hub/models/deepseek-v4-pro) lists $0.66 input and $1.98 output off-peak, and the [cheapest LLM API guide](/guides/cheapest-llm-api-2026) ranks the field.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Causal Encoder-Decoder Architecture: Runs a dedicated encoder stage ahead of the decoder stage, activating only a small fraction of the full Mixture-of-Experts backbone on each token instead of running every expert.
• 1,048,576-Token Context Window: Supports over 1 million tokens of context, with a documented API output cap of 384,000 tokens.
• Native Vision Encoder: A purpose-built DeepSeek-ViT vision encoder feeds image embeddings into the model jointly with text, rather than as a bolted-on adapter.
• Continuous Reasoning Effort Control: Accepts any integer reasoning_effort from 1 to 100, with API aliases low/high/max mapping to 50/75/100, letting callers trade latency for accuracy in fine steps.
• Peak/Off-Peak API Pricing: Halves every rate outside two weekday UTC windows, with Chinese public holidays billed fully off-peak; see pricing for exact per-token rates.
• Search and Fetch APIs: DeepSeek opened its in-house search engine to API users in October 2026 as separate per-request services billed apart from token usage.

Pros

• Posts a 3,471 Codeforces rating at launch, and DeepSeek reports the October snapshot lifted its coding-agent scores across the board.
• MIT-licensed weights on Hugging Face allow free commercial self-hosting and fine-tuning, unusual among models with frontier-adjacent coding scores.
• A very large context window and off-peak API pricing among the lowest for a vision-capable model at this benchmark tier.

Cons

• No published SWE-bench Verified, AIME 2025, ARC-AGI-2 or LMArena score for this release, making direct frontier comparisons hard.
• Text and image input only; no native audio or video modality.
• Artificial Analysis recorded 250M output tokens to run its index against a 140M median for comparable models, so the low per-token price overstates the cost advantage.
• No system card is linked from the model card or API docs, and Claude Opus 5 leads in the vendor-published October comparison table on several reasoning and coding rows.

Benchmarks

• MMLU-Pro: 74.1% vendor-reported · 13 Sep 2026 — A harder version of the 57-subject knowledge exam, % correct.
• HumanEval: 79.4% vendor-reported · 13 Sep 2026 — Small programs that must pass hidden tests, % passing.
• GPQA Diamond: 90.9% vendor-reported · 19 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
• Terminal-Bench 2.1: 90.6% vendor-reported · 19 Oct 2026 — Multi-step tasks completed in a real command line, % solved.
• Humanity's Last Exam: 63.9% vendor-reported · 19 Oct 2026 — Expert-written questions across many fields, % correct.
• AA Intelligence Index: 39 cited: Artificial Analysis · 19 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• AA blended price: $0.18/M cited: Artificial Analysis · 19 Oct 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
• Output speed: 231 tok/s cited: Artificial Analysis · 19 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit DeepSeek-V4.1-Flash Official Page