DeepSeek 4.1 Flash: A Third of the V4-Pro Price, With Vision
DeepSeek 4.1 Flash: $0.15 to $0.30 per million input tokens, 1M context, 2,500 concurrency, and the only DeepSeek model with vision. Pricing table and tips.

DeepSeek 4.1 Flash is the model you get when you send deepseek-flash to the API — version string DeepSeek-V4.1-Flash-1017. It is one of exactly two models DeepSeek currently sells, next to deepseek-v4-pro (DeepSeek-V4-Pro-0813), and on the official pricing page it is the cheaper one by a wide margin: $0.15 per million input tokens off-peak, $0.60 per million output.
It is also, quietly, the more capable one on one axis. Flash accepts images; V4-Pro does not. Add a 1M-token context, 384K max output and 2,500 concurrent connections — five times the 500 V4-Pro allows — and "Flash" starts to look less like the budget tier and more like the default.
This post covers the full pricing table (off-peak and peak, cache hit and miss), the Flash-vs-Pro decision, the silent migration that already moved your old deepseek-v4-flash calls, and the scheduling and caching habits that cut a real bill by half or more.
What DeepSeek 4.1 Flash is
Two facts explain most of what follows.
First, the lineup is tiny. The DeepSeek API offers deepseek-flash and deepseek-v4-pro, and that is the whole menu. Flash is DeepSeek-V4.1-Flash-1017; Pro is DeepSeek-V4-Pro-0813. Note the version strings: Flash carries the newer 4.1 generation, Pro is still V4, and the four-digit suffixes read as month-day release stamps — October 17 for Flash, August 13 for Pro.
Second, thinking mode is the default. Send a request and you get a reasoning model unless you turn thinking off. That matters for one feature in particular: fill-in-the-middle (FIM) completion exists only in non-thinking mode. If you are wiring Flash into an editor for inline completions, those calls need thinking disabled.
| Spec | DeepSeek 4.1 Flash |
|---|---|
| API model name | deepseek-flash |
| Version | DeepSeek-V4.1-Flash-1017 |
| Context window | 1M tokens |
| Max output | 384K tokens |
| Concurrency | 2,500 connections |
| Default mode | Thinking |
| Inputs | Text and images (vision) |
| Features | JSON output, tool calls, Responses API, Anthropic-compatible API |
| FIM completion | Non-thinking mode only |
DeepSeek 4.1 Flash pricing: all six numbers
| Per million tokens | Off-peak | Peak |
|---|---|---|
| Input, cache hit | $0.003 | $0.006 |
| Input, cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Off-peak is exactly half of peak across the board. And "peak" is narrower than you would guess: weekdays only, 01:00-04:00 and 06:00-10:00 UTC, excluding Chinese public holidays. In Beijing time that is 09:00-12:00 and 14:00-18:00 — office hours with a lunch break. Everything else, including every weekend, is off-peak.
For a US East Coast team, the entire working day falls in off-peak; the West Coast only brushes the first peak block at 5pm PST. For Europe, the morning block (07:00-11:00 CET) lands on the second peak window, so a pipeline that runs at 9am Berlin pays double what the same job costs at 1pm.
Cache hits are the other half of the story. A cache-hit input token costs $0.003 off-peak — 2% of the cache-miss rate, a 50x difference. Agent loops that resend the same system prompt and tool definitions every turn, RAG pipelines that prepend the same documents, eval harnesses that run one prompt across hundreds of cases: once warm, those are mostly hits, and their input bill is close to a rounding error.
A worked example: 10M input tokens, 80% cache hits, off-peak. That is 8M at $0.003 plus 2M at $0.15 — $0.024 plus $0.30, or $0.32 total. The same tokens with no caching: $1.50. No caching, at peak: $3.00. Same work, roughly a 10x spread, decided entirely by when you run and how you structure prompts.
Flash vs V4-Pro: price, vision, concurrency
| DeepSeek 4.1 Flash | DeepSeek V4-Pro | |
|---|---|---|
| API model name | deepseek-flash | deepseek-v4-pro |
| Version | DeepSeek-V4.1-Flash-1017 | DeepSeek-V4-Pro-0813 |
| Input, cache miss (off-peak) | $0.15 | $0.66 |
| Output (off-peak) | $0.60 | $1.98 |
| Vision input | Yes | No |
| Concurrency | 2,500 | 500 |
Do the division: off-peak to off-peak, Flash is 23% of V4-Pro on input and 30% on output. Call it a quarter to a third of the price — the V4-Pro premium works out to 4.4x on input and 3.3x on output.
Then the two non-price lines. Vision: Flash takes images, V4-Pro does not, full stop. If your DeepSeek pipeline needs screenshots, scanned documents or UI captures, Flash is not the cheap option — it is the only option. Concurrency: 2,500 connections against 500 means a Flash deployment can run five times as many simultaneous requests before anything queues.
What V4-Pro buys you is capability on the hardest tasks, if your evals confirm it — "Pro" is a name, not a guarantee. The pricing page already tells you the cost of finding out: run the same 50 prompts through both and compare.
Throughput: what 2,500 concurrent connections means for builders
Concurrency limits are the spec nobody reads until launch day. For a hosted API, 2,500 simultaneous connections is a lot — enough to fan out an agent swarm, batch-classify a backlog, or serve a consumer product without building a queue in front of DeepSeek. The 500 on V4-Pro is fine for an internal tool and a hard ceiling for anything user-facing at scale.
Put the concurrency next to 384K max output and the 1M window and Flash is the DeepSeek model for pipelines, not just chat: long document rewrites, multi-file code generation, extraction jobs that emit a lot of structured JSON.
What changed: the retired model names
If you are still sending deepseek-v4-flash, here is what is happening. The name is accepted, the model behind it is retired, and the request is silently served by DeepSeek-V4.1-Flash-1017 at Flash prices. Nothing errors. Your logs say one model; your outputs come from another.
That is convenient and slightly dangerous. Convenient because nothing broke. Dangerous because any eval baseline, prompt tuning or output-format assumption you built on the old model is now being tested against a different one, with no changelog entry in your repo.
What to do: update the model string to deepseek-flash, re-run whatever regression set you have, and confirm the thinking-mode default matches what your prompts expect (thinking is on by default; FIM needs it off). Check the official docs for the full list of legacy names rather than assuming only the flash alias moved.
Flash or V4-Pro: a short decision rule
Flash if any of these is true: you need image input; you run more than 500 concurrent requests; your volume is high enough that a 3-4x price difference shows up on an invoice; you want the newer 4.1-generation model.
V4-Pro if your own evals show it winning on tasks where Flash fails, you are text-only, and you stay under 500 concurrent connections.
Default: start on Flash, measure, and move specific hard tasks to Pro if the numbers justify it — not the other way around.
Practical tips for the bill
Schedule around UTC, not your local clock. A nightly batch job set for 02:00 US Eastern runs at 07:00 UTC — inside the second peak block. Shift it to 11:00 UTC or later, or to the weekend, and the same job is half price.
Build prompts cache-first. Stable prefix (system prompt, tool schemas, reference documents) at the top, variable content at the bottom. Hits at $0.003 versus misses at $0.15 is the largest ratio on the pricing page.
Turn thinking off where you do not need it. Thinking mode is the default; it spends tokens on reasoning before the answer, and output is the expensive column in the table. Classification, extraction, FIM and short-form generation usually do not need it — switch it off and measure the difference.
Budget the Search and Fetch APIs separately. The DeepSeek Search and Fetch endpoints cost $0.75 per 1,000 successful requests each. The free period ended when billing started on November 1, 2026 (Beijing time) — if you built on them during the free window, that line item just appeared.
Who DeepSeek 4.1 Flash fits
Teams with volume. Anyone building agents that fan out. Anyone who needs image input from a DeepSeek model, since there is no other option. And, honestly, most people who were about to default to V4-Pro because "Pro" sounds safer.
If you are weighing it against the rest of the open-model field: Kimi K3.1 Pro is the long-context play at $2/$10 — 13x the Flash off-peak input price, with open weights due November 4 — and GLM-5.3 is the third name in the current cycle. For setup, our guide on how to use DeepSeek V4.1 walks through the API, and the DeepSeek listing tracks the product itself. If you want Flash inside a terminal agent, the Anthropic-compatible endpoint is the route that tools built around Claude Code-style workflows tend to use — check the docs for which features survive the compatibility layer. Our AI coding tools like Cursor roundup covers the agents themselves.
FAQ
What is the difference between deepseek-flash and DeepSeek V4.1 Flash?
Same model. deepseek-flash is the API model name; DeepSeek-V4.1-Flash-1017 is the version string behind it. Legacy names such as deepseek-v4-flash are also routed to it at Flash prices.
How much does DeepSeek 4.1 Flash cost?
Per million tokens: cache-miss input $0.15 off-peak / $0.30 peak, cache-hit input $0.003 / $0.006, output $0.60 / $1.20. Peak is weekdays 01:00-04:00 and 06:00-10:00 UTC (excluding Chinese public holidays); everything else is off-peak at half price.
Does DeepSeek V4-Pro support vision?
No. In the current two-model lineup, only DeepSeek 4.1 Flash accepts image input. V4-Pro is text-only, costs $0.66 input / $1.98 output off-peak, and allows 500 concurrent connections against 2,500 on Flash.
Is thinking mode on by default?
Yes. Flash defaults to thinking mode. Fill-in-the-middle completion is only available with thinking off, so editor-style completion calls need to disable it explicitly.
How big is the DeepSeek 4.1 Flash context window?
1M tokens, with up to 384K tokens of output per request. That is enough for whole-repository or long-document tasks in one call — just remember that a full 1M-token input is about $0.15 to $0.30 of cache-miss tokens before any output.


