Skip to content
Models / DeepSeek

DeepSeek-V4.1-Flash

DeepSeek · Model release

Released Not recorded

Catalog checked 2026-09-28

Input $/1M
$0.15
Output $/1M
$0.60
Context
1M
Max output
384K
Reasoning
Yes
Included in
8 plans
API priceUSD / 1M tokens
Published API rates in US dollars per million tokens
RateInputOutputCache read
StandardInput$0.15Output$0.60Cache read$0.003
Peak: 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, except Chinese public holidaysInput$0.30Output$1.20Cache read$0.006

Published rate checked 2026-09-29. Sources below.

Current published base rates. Context tiers, cache-write duration and other conditions may change the rate for a request. These are not a reconstructed historical invoice.

API token estimate

Token estimate unavailable

This model has conditional rates. Aggregate token totals cannot establish which request or schedule conditions apply.

Capabilities & limitsProvider specifications
Context window
1,000,000
Maximum output
384,000
Input
text · image
Output
text
Reasoning
Supported
Tool calling
Supported
Structured output
Supported
Before you chooseChecked 2026-09-29
Thinking
Thinking is enabled by default; non-thinking mode is also supported.
Time-based pricing
Peak: 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. Other hours, weekends and those holidays use half-price off-peak rates.
API compatibility
OpenAI-compatible and Anthropic-compatible endpoints, Responses API, tool calls and JSON output.
Vision
Image input is supported.

DeepSeek · Published Sep 10, 2026

Developer reported · V4.1 Flash release

Show all 20 results
Methodology & sources

DeepSeek’s September 10 changelog reports these exact benchmark versions and scores. Configuration notes from older releases are not assumed to apply to V4.1 Flash.

These scores are reported by DeepSeek. Model developers identify authorship; results may originate with another evaluator. Exact benchmark versions remain distinct. StackReplay calculates no composite score.

DeepSWE v1.1 · 74.2%

Long-horizon software engineering in real-world codebases.

Version: v1.1 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not reported for this release
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

DeepSeek’s September 10 changelog reports these exact benchmark versions and scores. Configuration notes from older releases are not assumed to apply to V4.1 Flash.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Terminal-Bench 4.0 · 31.2%

Agentic task completion in a terminal environment.

Version: 4.0 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not reported for this release
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

DeepSeek’s September 10 changelog reports these exact benchmark versions and scores. Configuration notes from older releases are not assumed to apply to V4.1 Flash.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Codeforces Rating · 3471

Competitive programming performance reported as a rating.

Version: Not reported · Rating · points · Higher is better · Task subset: Rating

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Competitive programming performance reported as a rating. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Terminal-Bench 2.1 · 90.6%

Multi-step tasks in a command-line environment.

Version: 2.1 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Multi-step tasks in a command-line environment. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Terminal-Bench 3.0 · 30.0%

Multi-step tasks in a command-line environment.

Version: 3.0 · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Multi-step tasks in a command-line environment. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

ProgramBench · 20.3%

Program construction evaluated by the benchmark’s tests.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Program construction evaluated by the benchmark’s tests. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

NL2Repo-Bench · 65.4%

Building software repositories from natural-language requirements.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Building software repositories from natural-language requirements. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Automation-Bench DeepSeek release; subset unspecified · 54.8%

Workflow automation in the evaluation named by DeepSeek.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: DeepSeek release; subset unspecified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Workflow automation in the evaluation named by DeepSeek. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Agents’ Last Exam · 31.8%

Extended agent tasks combining reasoning and execution.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Extended agent tasks combining reasoning and execution. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

GPQA Diamond · 90.9%

Graduate-level scientific question answering.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Graduate-level scientific question answering. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Humanity’s Last Exam Without tools · 36.8%

Expert-level multidisciplinary questions without tools.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Without tools

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Expert-level multidisciplinary questions without tools. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Humanity’s Last Exam Pure-text subset · 39.1%

The pure-text subset of expert-level multidisciplinary questions.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Pure-text subset

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

The pure-text subset of expert-level multidisciplinary questions. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

MathArena Apex · 65.6%

Advanced mathematical problem solving.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Advanced mathematical problem solving. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Humanity’s Last Exam With tools · 63.9%

Expert-level multidisciplinary questions with tools.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: With tools

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Expert-level multidisciplinary questions with tools. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

BabyVision With tools · 89.6%

Visual reasoning evaluated with tool access.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: With tools

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Visual reasoning evaluated with tool access. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

ZeroBench main With tools · 49.0%

Challenging multimodal reasoning with tools.

Version: main · Reported benchmark score · percent · Higher is better · Task subset: With tools

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Challenging multimodal reasoning with tools. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Chartography · 78.9%

Visual interpretation and analysis of charts.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result
Harness
Not reported
Tools
With tools
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

DeepSeek reports Chartography with tools. Google and Anthropic results without tools use different configurations.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

CyberGym · 88.1%

Agentic cybersecurity tasks in realistic software environments.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Agentic cybersecurity tasks in realistic software environments. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

SEC-Bench Pro · 62.8%

Professional software security tasks.

Version: Pro · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Professional software security tasks. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

ExploitGym · 15.3%

Security exploit development tasks.

Version: Not reported · Reported benchmark score · percent · Higher is better · Task subset: Not separately specified

DeepSeek · Published Sep 10, 2026 · Developer reported

Reporting source
DeepSeek · V4.1 Flash release
Evaluation origin
Not separately specified in the cited release summary (evaluation runner not reported)
Effort
Not specified for this result in the release summary
Harness
Not reported
Tools
Not reported
Fallback
Not reported
Provider
Not reported
Checked
Sep 30, 2026

Security exploit development tasks. The score is transcribed from the official release summary; unreported configuration is not inferred.

Harness, task subset and statistical uncertainty are not fully specified in the cited announcement.

Original methodology ↗ · Original evidence ↗

Where you can use it1 API route · 8 subscriptions
DeepSeek APIDirect API
ClinePass
$9.99 / month ↗

$9.99 monthly; weekly and monthly usage windows follow the calendar.

Command Code Go
$1 / month ↗

Monthly cycle; purchased credits carry over without expiry.

New subscriptions use per-model allowances from September 28, 2026. Existing flat-$10 subscriptions change at their next renewal.

Command Code Pro
$20 / month ↗

Monthly cycle; purchased credits carry over without expiry.

Command Code Max 10×
$100 / month ↗

Monthly cycle; purchased credits carry over without expiry.

Command Code Max 20×
$200 / month ↗

Monthly cycle; purchased credits carry over without expiry.

Ollama Cloud Max
$100 / month ↗

Monthly credit anniversary; included credits do not carry over.

OpenCode Go
$10 / month ↗

Monthly subscription; one subscriber per workspace. Paid balance fallback is optional. Recurring charges continue until canceled through account settings. Reset anchors, rollover and a Go-specific refund policy are not established by the reviewed documentation.

OpenCode Go Plus
$40 / month ↗

Monthly subscription; one subscriber per workspace. Paid balance fallback is optional. Recurring charges continue until canceled through account settings. Reset anchors, rollover and a Go-specific refund policy are not established by the reviewed documentation.

Pricing, assumptions & evidence

Published base rate

Input $0.15 · Output $0.60 · Cache read $0.003 · Cache write Not listed

Peak: 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, except Chinese public holidays: input $0.30, output $1.20, cache read $0.006.

Missing token-category prices are not zero. Workload pricing applies exact recorded categories and admitted routes. No benchmark score or quality ranking is inferred from prices.

Aliases, routes and identity

Catalog ID deepseek-v4-1-flash

Model release (kind: release)

Developer: DeepSeek

deepseek-flash

provider_id