Model comparisons
DeepSeek V4.1 Flash vs V4 Pro: Pro Costs 3-7x More and Trails on Agent Benchmarks
DeepSeek's deepseek-flash (V4.1-Flash-1017) costs $0.30/$1.20 per 1M tokens at peak against deepseek-v4-pro's $1.32/$3.96. In DeepSeek's September model card, V4.1 Flash beats V4 Pro on Terminal-Bench 4.0 (31.2% vs 12.4%) and DeepSWE (74.2% vs 62.7%); Pro leads on GPQA Diamond and text-only HLE.
Updated
In short
Use deepseek-flash, DeepSeek's V4.1 Flash, for nearly all work, coding agents included. At peak hours it costs $0.30 per million input tokens and $1.20 per million output tokens against deepseek-v4-pro's $1.32 and $3.96, so Pro costs 4.4 times as much per uncached input token, 3.3 times as much per output token and 7.3 times as much per cached input token. In the instruct comparison DeepSeek published with V4.1 Flash, Flash also beats V4 Pro on 14 of 16 rows, by the most on Terminal-Bench 4.0 (31.2% vs 12.4%). Pro is ahead there only on GPQA Diamond (92.4% vs 90.9%) and tool-free Humanity's Last Exam, and it reads no images. Both are API models with a 1M-token context and 384K maximum output.
On this page5
V4.1 Flash beats V4 Pro on 14 of 16 rows in DeepSeek's own table
DeepSeek's V4.1 Flash model card (September 10, 2026) runs both models at maximum reasoning effort in the same table, with 16 rows that score both. The first two columns below are DeepSeek's figures for ten of those rows. The third column is from the model card for V4.1-Flash-1017, the October 17 update now served as deepseek-flash, which reports only some of the same rows and does not rerun V4 Pro.
Benchmark (DeepSeek model cards) | V4 Pro | V4.1 Flash (September) | V4.1-Flash-1017 (current) |
|---|---|---|---|
GPQA Diamond | 92.4% | 90.9% | not reported |
Humanity's Last Exam, text-only, no tools | 42.7% | 39.1% | 44.0% |
Humanity's Last Exam, with tools | 60.0% | 63.9% | not reported |
Codeforces rating | 3348 | 3471 | not reported |
Terminal-Bench 2.1 | 87.9% | 90.6% | not reported |
Terminal-Bench 4.0 | 12.4% | 31.2% | 42.3% |
DeepSWE v1.1 | 62.7% | 74.2% | 77.0% |
SEC-Bench Pro | 56.4% | 62.8% | not reported |
AutomationBench | 43.2% | 54.8% | 60.5% |
Agent's Last Exam | 25.7% | 31.8% | 35.4% |
Read the agent rows first: the gap is widest where a model has to work in a terminal or repository for many steps, which is what a coding agent bills you for. The September card runs Terminal-Bench in DeepSeek Harness's Minimal mode, DeepSWE in mini-SWE and SEC-Bench Pro in Claude Code; the October card runs Terminal-Bench 4.0 in Claude Code and DeepSWE in Mini-SWE-V2, so set V4 Pro only against the September column. Both cards are DeepSeek's own evaluations.
The same September card also compares the base models before tuning, and there V4 Pro still leads on factual recall (SimpleQA-Verified 55.2 vs 42.3) and long context (LongBench-V2 51.5 vs 45.2). Those are the workloads where Pro may be worth measuring against Flash.
Pro costs 3.3 to 7.3 times as much, and off-peak halves both
DeepSeek bills both models per million tokens from a prepaid balance, with no plans or tiers. Off-peak rates are exactly half of peak rates for both models, so the ratio between them does not change with the clock.
Per 1M tokens | deepseek-flash, peak | deepseek-v4-pro, peak | deepseek-flash, off-peak | deepseek-v4-pro, off-peak | Pro costs |
|---|---|---|---|---|---|
Input, cache hit | $0.006 | $0.044 | $0.003 | $0.022 | 7.3x |
Input, cache miss | $0.30 | $1.32 | $0.15 | $0.66 | 4.4x |
Output | $1.20 | $3.96 | $0.60 | $1.98 | 3.3x |
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays; weekends and those holidays are off-peak all day. For a team in the Americas the whole working day falls off-peak, while European mornings and Asia-Pacific working hours overlap the peak windows.
The cache-hit row matters most for agents. DeepSeek itself says cache-hit charges often account for a large share of agent costs: a coding agent resends its whole conversation on every turn, so most of its input tokens are cache hits, and that is where Pro's premium is largest. In ToolColumn arithmetic, an illustrative agent turn that rereads 70,000 cached tokens, adds 400 new input tokens and writes 300 output tokens costs about $0.0009 on deepseek-flash and $0.0048 on deepseek-v4-pro at peak. A thousand such turns cost about $0.90 against $4.80, or $0.45 against $2.40 off-peak, a gap of about 5.3 times.
Flash reads images and allows five times the concurrency
API specification | deepseek-flash | deepseek-v4-pro |
|---|---|---|
Model version | DeepSeek-V4.1-Flash-1017 | DeepSeek-V4-Pro-0813 |
Context / maximum output | 1M / 384K tokens | 1M / 384K tokens |
Image input | yes | no |
Concurrent requests per account | 2,500 | 500 |
Thinking mode | on by default, effort low, high or max | on by default, effort low, high or max |
Formats | OpenAI, Anthropic, Responses | OpenAI, Anthropic, Responses |
DeepSeek counts concurrency per account across all API keys and returns HTTP 429 above it; you can request more at no extra cost.
Claude Code users can land on Pro without choosing it; Codex search needs it
DeepSeek's Anthropic-format endpoint translates Claude model names. Any name starting with claude-opus runs on deepseek-v4-pro at Pro prices, and claude-sonnet or claude-haiku names run on deepseek-flash. DeepSeek's own Claude Code setup points every model slot at deepseek-flash, so an agent configured that way never touches Pro; one that leaves an Opus model selected pays Pro rates for it.
Codex is the reverse case. DeepSeek's Codex setup says Codex's built-in web search works only with deepseek-v4-pro, and those searches have been billed at $0.75 per 1,000 successful requests since 00:00 Beijing time on November 1, 2026.
Pro stays on sale despite the September phase-out notice
DeepSeek's September 10 announcement said that from September 14, 2026 all deepseek-v4-pro requests would route to V4.1 Flash at Flash rates until a V4.1 Pro launches. Its API change log entry for the same date says the opposite: V4 Pro service continues after September 14 with billing unchanged. The pricing page, read on October 31, 2026, lists DeepSeek-V4-Pro-0813 at its own prices, which matches the change log. DeepSeek has not announced a V4.1 Pro.
The free app does not let you choose: DeepSeek's change log says the web and app moved to V4.1-Flash-1017 on October 17, its iOS release notes of September 11 say the Instant, Expert and Vision modes were unified, and DeepSeek names no way to reach V4 Pro there.
Edited byRick Finch
Official sources
Editorial guidance grounded in official product sources.
- DeepSeek API models and pricing
- DeepSeek API change log
- Introducing DeepSeek-V4.1-Flash (DeepSeek, September 10, 2026)
- DeepSeek-V4.1-Flash model card (Hugging Face)
- DeepSeek-V4.1-Flash-1017 model card (Hugging Face)
- DeepSeek Anthropic API guide
- DeepSeek rate limit and isolation
- DeepSeek Codex integration
Show all 10 sources
FAQ
Common questions
DeepSeek V4.1 Flash vs V4 Pro, which should I use?
Use deepseek-flash for almost everything, coding agents included. Pro costs 3.3 times as much per output token, 4.4 times per uncached input token and 7.3 times per cached input token, and Flash leads it on 14 of 16 rows of the instruct comparison in DeepSeek's own model card. Keep deepseek-v4-pro for work you have measured to do better on it, such as factual recall, where the card's base-model table still favors Pro.
How much does DeepSeek V4 Pro cost?
At peak hours deepseek-v4-pro costs $1.32 per 1M input tokens on a cache miss, $0.044 on a cache hit and $3.96 per 1M output tokens. Off-peak, outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, it costs half: $0.66, $0.022 and $1.98.
Is DeepSeek V4 Pro being discontinued?
Not as of October 31, 2026. DeepSeek's September 10 announcement said deepseek-v4-pro requests would route to V4.1 Flash from September 14 until a V4.1 Pro launches, but its API change log says V4 Pro service continues with billing unchanged, and the pricing page still lists DeepSeek-V4-Pro-0813 at its own prices.
Which DeepSeek model does Claude Code use?
It depends on the model name Claude Code sends to DeepSeek's Anthropic endpoint. Names starting with claude-opus run on deepseek-v4-pro and are billed at Pro prices; claude-sonnet and claude-haiku names run on deepseek-flash. DeepSeek's own Claude Code setup sets every model slot to deepseek-flash.
Does DeepSeek V4 Pro support images?
No. DeepSeek's pricing page marks vision as not supported on deepseek-v4-pro, while deepseek-flash accepts images. The legacy deepseek-v4-flash-vision-exp name now routes to deepseek-flash.
What is the difference between deepseek-flash and deepseek-v4-flash?
deepseek-flash is the current API name and runs DeepSeek-V4.1-Flash-1017. The old V4 Flash model is retired; for now DeepSeek still accepts the deepseek-v4-flash name but serves it with V4.1-Flash-1017 at the deepseek-flash price.
Next steps
Open both sides of the distinction
- Tool profileReview DeepSeek pricingSee DeepSeek's free app, both API rate cards and the Search and Fetch fees on one page.
- Tool profileOpenCodeRun both DeepSeek models in a coding agent with a fixed monthly allowance instead of a token balance.
- GuideClaude Haiku 5.5 vs GPT-6 Luna: Same Price, Different FitCompare DeepSeek's prices with Claude Haiku 5.5 and GPT-6 Luna, the small models from Anthropic and OpenAI, before choosing one for high-volume work.