DeepSeek V4.1 Flash 1017: Benchmarks, Pricing and Setup

What Is DeepSeek V4.1 Flash 1017?

DeepSeek-V4.1-Flash-1017 is the newest version of DeepSeek's V4.1-Flash model, and it is live on DeepSeek's website, app and API. The release note calls it a Pareto improvement in performance and cost, and says the model is optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy.

DeepSeek announced the release on X:

The update builds on the architecture DeepSeek introduced on September 10, 2026. Its V4.1-Flash release note describes a 552B-parameter mixture-of-experts model with a causal encoder-decoder design that activates 8B parameters for input and 16B for output, and a KV cache that needs a quarter of the HBM and an eighth of the SSD storage of the previous generation.

DeepSeek's models and pricing page lists what the API serves under the deepseek-flash ID:

How Does DeepSeek V4.1 Flash 1017 Score on Benchmarks?

DeepSeek's model card compares DeepSeek-V4.1-Flash-1017 with DeepSeek-V4.1-Flash and four other models: K3, GLM-5.3, GPT-5.6 Sol and Opus 5.

K3 is Kimi K3 and Opus 5 is Claude Opus 5, as the version of the table in DeepSeek's post on X labels them. The scores are vendor-reported. In the table below, the September column is the previous V4.1-Flash checkpoint, and the last column is the best result among those four other models.

The notes under the table and in the change log explain how the runs were set up, which decides how far the numbers carry over to your agent:

• Harnesses - Terminal-Bench 4.0 ran in the Claude Code harness and DeepSWE v1.1 in the Mini-SWE-V2 harness, so each score measures the model inside that harness, not the model on its own.
• HLE subset - The bracketed HLE figures are the text-only subset of the benchmark.
• ZeroBench-main - This row is scored Pass@5, so the model gets five attempts per item.
• Comparison set - The table stops at Claude Opus 5 and GPT-5.6 Sol, and leaves out newer models such as Claude Opus 5.5.

Which rows matter depends on the work your agent does:

• Issue fixing and automations - The 1017 checkpoint posts the top score in the table on DeepSWE v1.1, Automation-Bench v1.0.6, Agents' Last Exam and Apex Bench v1.1, the rows closest to an agent working through tickets and multi-step workflows.
• Terminal and from-scratch repository work - Claude Opus 5 leads Terminal-Bench 4.0 by 9.5 points, ProgramBench by 7.9 and NL2Repo-Bench by 7.1, so long shell sessions and building a repository from a spec still favor it. For a wider shortlist, see the best LLM for coding.
• Reasoning without tools - The widest gap in the table is HLE, where Claude Opus 5 scores 13.7 points higher.
• Visual tasks - Claude Opus 5 leads Chartography by 1.0 point and BabyVision by 1.6, and GPT-5.6 Sol leads ZeroBench-main by 2.0.
• Gains over September - Terminal-Bench 4.0 rose 11.1 points and ProgramBench 8.8, the two largest jumps, so terminal-heavy agents are where a retest is most likely to show a change.
• Cost per task - The release note adds a DeepSWE v1.1 chart of Pass@1 against average cost per task, which plots the 1017 checkpoint's max-effort run above and to the left of the September checkpoint's max-effort run: a higher score at a lower cost.

Note: DeepSeek's scores come from its own harness runs, not from your prompts, tools or users. TestMu AI Agent Testing generates test scenarios from your agent's own documents, including adversarial and persona-specific cases, so you can compare checkpoints on the conversations your agent actually handles. Create a free account

How Much Does DeepSeek V4.1 Flash 1017 Cost on the API?

DeepSeek bills deepseek-flash per million tokens, at peak and off-peak rates, and the 1017 update left those rates unchanged. The models and pricing page lists them next to deepseek-v4-pro:

• Peak hours - 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Every other hour, including weekends and those holidays in full, bills at the off-peak rate, which is half the peak rate.
• Cache hits - A cache-hit input token costs one-fiftieth of a cache-miss token on deepseek-flash, so an agent that resends the same system prompt and tool definitions every turn pays mostly the cache-hit rate on that repeated prefix.
• Legacy model names - Requests that still name deepseek-v4-flash or deepseek-v4-flash-vision-exp are served by the 1017 checkpoint and billed at the Flash price.
• Scheduling - Nightly evaluation suites and batch jobs that can wait should run off-peak, where every rate in the table is halved.

How the DeepSeek API Handles Thinking, Effort and Model Names

Several request settings behave differently from what OpenAI- and Anthropic-style client code assumes.

The thinking mode guide documents the thinking defaults, how requested effort values map to the effort the model actually uses, and which sampling parameters take effect:

On the Anthropic API endpoint, the model name you send decides which DeepSeek model runs and what you pay:

• claude-opus names - Map to deepseek-v4-pro and are billed at the V4 Pro price, not the Flash price.
• claude-sonnet and claude-haiku names - Map to deepseek-flash, which now means the 1017 checkpoint.
• Other unsupported names - A model name that is not a DeepSeek model and does not start with claude returns a 400 error.
• Ignored fields - Thinking budget_tokens, cache_control, mcp_servers, top_k and the anthropic-version header are accepted and ignored, and output_config supports only effort.

Check two things in your client before anything else: the effort value it sends, because xhigh runs at high, and any test that sets temperature to 0 expecting a repeatable answer, because thinking mode ignores it.

How to Regression Test an Agent After the 1017 Update

A checkpoint swap under the same model ID is a change you did not deploy, so treat it like one: run the agent's existing suite against the new checkpoint and compare the results with the last run from before October 17. Agentic regression testing covers which of those checks to automate, and LLM testing covers the basics of scoring model output. Each row below traces to a line in DeepSeek's documentation:

Checking Chat Agents for Shorter Answers

Shorter outputs lower a coding agent's token bill, and they can also cut a support agent's answer short. TestMu AI's Agent Testing is built for regression testing after model updates: it runs a chat agent's scenario suite through the same endpoint your users reach and scores every reply, so a run on the 1017 checkpoint can be set against the run before it.

• Completeness - Checks whether a reply fully addresses the user's need. A user who asks how to cancel and gets the cancellation policy without the steps fails it, which is the failure shorter outputs make more likely.
• Hallucination and context awareness - Flags invented facts and details the agent forgets from earlier turns, two behaviors any checkpoint change can move.
• Go or no-go verdict - Each run ends in Green, Yellow or Red with a confidence level, so the model swap gets a release decision instead of a folder of transcripts.
• Endpoint profiles - A saved REST, WebSocket, SSE or polling profile defines how the platform reaches your agent, so the same suite reruns unchanged whenever the model underneath it changes.

How to Use DeepSeek V4.1 Flash 1017 in a Coding Agent

The release note says the 1017 checkpoint uses reinforcement learning with self-feedback to improve end-to-end project development, and shows it building a web-based racing game, a 3D model of the Forbidden City and a Portal 2 recreation in the Godot engine. It recommends DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.

DeepSeek Harness, DeepSeek's own agent harness, added support for DeepSeek-V4.1-Flash-1017 in dsh-v0.2.1-rc.2, along with the DeepSeek Search and Fetch APIs, which you turn on in its Web Search settings.

For Claude Code, DeepSeek's integration guide sets ANTHROPIC_BASE_URL to DeepSeek's Anthropic endpoint, sets ANTHROPIC_MODEL and the default Opus and Sonnet models to deepseek-flash[1m], and sets the Haiku and subagent models to deepseek-flash:

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=<your DeepSeek API key>
export ANTHROPIC_MODEL=deepseek-flash[1m]
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-flash[1m]
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-flash[1m]
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-flash
export CLAUDE_CODE_EFFORT_LEVEL=max
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=786432

Keep the ANTHROPIC_DEFAULT_OPUS_MODEL line. Without it, Claude Code's Opus slot sends a claude-opus model name, which DeepSeek routes to deepseek-v4-pro at the V4 Pro price.

DeepSeek's demos are browser games and 3D scenes, and a coding agent that builds one can compile it, lint it and pass its unit tests without ever opening the page. Kane CLI runs a plain-English objective in a real Chrome browser and returns an evidence-backed pass or fail, so the agent can check the rendered result before it reports the task done:

npm install -g @testmuai/kane-cli
npx @testmuai/kane-cli-skill
kane-cli run "Open http://localhost:5173, click Start Race, and verify that the lap timer starts counting and the car is visible on the track" --agent --headless

The skill installer adds Kane CLI to Claude Code, Codex CLI and Gemini CLI in one command, and the guide to Kane CLI with AI coding agents covers the other builders. With --agent, every run streams one JSON event per line and ends with a run_end event holding the status, summary and log paths, plus an evidence pack of screenshots, video and console logs the coding agent can read when a step fails.

How to Get Started With DeepSeek V4.1 Flash 1017

Pull your agent's last passing run from before October 17, rerun the same scenarios against deepseek-flash at the effort value you ship, and compare the two before you send the agent new traffic. If replies keep their length and step counts look right, schedule the full suite for off-peak hours. The Agent Testing CLI docs show how to start that suite from CI with one command after every model change.