DeepSeek V4.1 Flash (2026): Price, Benchmarks, Real Cost

Short answer: DeepSeek shipped V4.1 Flash on September 10, 2026 at $0.15 per million input tokens and $0.60 per million output, with cache hits at $0.003. It is a 552B-parameter mixture-of-experts model that activates 8B parameters during prefill and 16B during decode, with a 1M-token context, native image input, and MIT-licensed weights. A 60-turn coding session with an 80% cache hit rate costs about $0.13 on it, against $4.42 on GPT-5.6 Sol and $5.52 on Claude Opus 5. From September 14, DeepSeek routes every V4 Pro request to it, whether or not you ask.

DeepSeek did three things at once on September 10: released a new model, cut its flagship price by roughly 60% to 70%, and started the countdown on retiring the model it just undercut. The third one reaches people who never opened a DeepSeek account, since it changes what deepseek-v4-pro does in code they already run.

What follows is what shipped, what it charges, and what a week of agent work costs.

DeepSeek V4.1 Flash: what shipped on September 10

V4.1 Flash is the small tier of a new architecture family, and DeepSeek's framing is blunt: it beats V4 Pro on performance, cost, speed, and task runtime. Tests by "multiple parties" are cited as the reason for phasing V4 Pro out.

The architecture is where the price comes from. It is a 552B-parameter MoE with a Causal Encoder-Decoder layout: 40 layers split into a 20-layer encoder and a 20-layer decoder. The decoder's global KV cache is projected from the encoder's final hidden states instead of built layer by layer, letting the model run 8B active parameters per token on input and 16B on output. DeepSeek reports the KV cache at 890 bytes per token, a quarter of V4 Flash's footprint and 1/437th of V1's.

• 1M-token context window, maximum output 384K tokens
• Native vision. V4 Pro does not support images at all
• Reasoning effort as an integer from 1 to 100, trading cost for accuracy
• MIT-licensed weights on Hugging Face with a roughly 50-page technical report
• No Jinja chat template. The release ships a Python reference encoder plus deepseek-recipe, Rust libraries with Python bindings, so you write the prompt encoding yourself

That last point matters if you self-host: this is not a drop-in for a llama.cpp style stack, and deployments run to a couple of thousand GPUs.

The price, and the peak-hour trap

DeepSeek kept peak/off-peak pricing and made the gap big enough to plan around. Prices are per 1M tokens:

	DeepSeek V4.1 Flash off-peak	V4.1 Flash peak	DeepSeek V4 Pro off-peak
Input, cache hit	$0.003	$0.006	$0.022
Input, cache miss	$0.15	$0.30	$0.66
Output	$0.60	$1.20	$1.98

Off-peak is exactly half of peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday: 7 hours a day, 35 hours a week. Everything else, weekends included, bills at the lower rate.

That schedule suits Europe. In Amsterdam time, peak lands at 03:00 to 06:00 and 08:00 to 12:00 local, so your afternoon and evening run off-peak and a nightly batch job pays half price. From US Pacific, peak covers 18:00 to 21:00 and 23:00 to 03:00 instead.

The cache-hit number is the one to pause on. Agent workloads are mostly cached input, and $0.003 per million tokens is cheaper than moving that text across the network at all, as a commenter noted on Hacker News at launch. Against GPT-5.6 Sol's $0.40 cache read or Claude Opus 5's $0.50, the input side of a long agent loop stops being the expensive part.

What the benchmarks say, including the rows DeepSeek's summary skips

Every number below comes from DeepSeek's own tables at maximum reasoning effort. Treat them as vendor results: Artificial Analysis had not published an independent V4.1 Flash measurement as of launch day.

Benchmark	V4.1 Flash	V4 Pro	GPT-5.6 Sol	Claude Opus 5
GPQA Diamond	90.9	92.4	94.1	93.4
HLE	36.8	42.7	44.5	56.3
Terminal-Bench 2.1	90.6	87.9	88.8	89.1
Terminal-Bench 3.0	30.0	11.8	34.4	43.3
Terminal-Bench 4.0	31.2	12.4	39.9	51.8
DeepSWE v1.1	74.2	62.7	73.0	74.0
NL2Repo-Bench	64.0	61.5	56.8	75.3
ProgramBench	20.3	15.5	23.0	37.0
CyberGym	88.1	83.3	84.5	not reported
Agent's Last Exam	31.8	25.7	26.7	28.6
AutomationBench	54.8	43.2	45.8	50.3
MathArena Apex	65.6	65.3	not reported	65.3
Codeforces rating	3471	3348	not reported	not reported

Read the whole table and the story is mixed, not triumphant. V4.1 Flash beats GPT-5.6 Sol and Claude Opus 5 on the lightest agentic rows (Terminal-Bench 2.1, DeepSWE, Agent's Last Exam, AutomationBench) and is far ahead of the model it replaces. It also loses badly on the harder ones: Opus 5 scores 43.3 against 30.0 on Terminal-Bench 3.0, 51.8 against 31.2 on 4.0, and HLE at 36.8 against 56.3 is not a rounding difference.

Scores move with the harness too. On DeepSWE v1.1 the same model scores 74.2 under DeepSeek's minimal harness and 69.8 under Claude Code, so cross-lab comparisons partly compare scaffolds.

What a coding session actually costs

Model this: 60 turns, 30,000 input tokens per turn as context grows, 2,000 output tokens per turn, 80% of input cached. That is 1.8M input and 120K output tokens per session.

Model	One session	20 sessions/month	Multiple of V4.1 Flash
DeepSeek V4.1 Flash (off-peak)	$0.13	$2.61	1.0×
DeepSeek V4.1 Flash (peak)	$0.26	$5.21	2.0×
DeepSeek V4 Pro (off-peak)	$0.51	$10.14	3.9×
Gemini 3.8 Flash	$0.83	$16.56	6.4×
Claude Sonnet 5	$2.21	$44.16	16.9×
GPT-5.6 Terra	$2.45	$48.96	18.8×
Gemini 3.1 Pro	$2.45	$48.96	18.8×
GPT-5.6 Sol (promo)	$4.42	$88.32	33.9×
Claude Opus 5	$5.52	$110.40	42.4×
Claude Fable 5.1	$9.96	$199.20	76.4×

Rates come from DeepSeek, OpenAI, Anthropic, and Google on September 10, 2026. GPT-5.6 Sol's row uses promotional $4/$20 rates that run into November 2026; list is $5/$30. If your loop is cache-poor instead, resending fresh context every turn, the gap narrows: at a 0% hit rate V4.1 Flash costs $0.34 per session and Opus 5 costs $12.00. Still 35× cheaper, though the absolute numbers stop being trivial.

One dollar buys about 1.67M output tokens on V4.1 Flash off-peak, 0.10M on Sonnet 5, and 0.04M on Opus 5. To run the same math with your own token counts, our breakdown of a real monthly API bill uses this method, and the live calculator takes your assumptions instead of mine.

The part DeepSeek did not lead with

Two sentences into the pricing footnotes, the retirement plan appears: from 12:00 Beijing time on September 14, 2026, every request to deepseek-v4-pro routes to V4.1 Flash and bills at Flash rates, until V4.1 Pro ships at some undetermined date.

DeepSeek's notes call it an orderly phase-out. Developers on Hacker News and finance coverage read it differently, and the complaint is operational rather than sentimental. If you pinned deepseek-v4-pro because you measured it against your evals, a reroute changes your system without a version bump, a deprecation window, or an opt-out. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp got the same treatment, and they only "temporarily" route to the new model.

If V4 Pro sits in a pipeline where output drift costs money, run your eval set against V4.1 Flash before September 14 so the change is your finding, not a surprise in a diff review. DeepSeek also raised the concurrency ceiling to 2,500 from 500.

Using it from a chat app with your own key

Two clean paths, and both work in any client that accepts a custom endpoint.

Through OpenRouter. V4.1 Flash went live on release day, served by four providers: DeepSeek, NovitaAI, DeepInfra, and Venice. OpenRouter reports 142 tokens per second median throughput, 1.59 seconds round-trip latency, and full uptime over three days. If you have a key already, our OpenRouter key walkthrough takes a few minutes, and what the 5.5% credit fee does to your effective rate matters before you push real volume through it.

Direct through DeepSeek. The API is OpenAI-compatible at https://api.deepseek.com, with an Anthropic-format endpoint at https://api.deepseek.com/anthropic. Point a custom endpoint at it and use deepseek-flash as the model name. The custom endpoint guide covers the fields, and this explainer on compatible APIs is the background if you have not wired a base URL into an app before.

One honest note: DeepSeek is not a BYOK provider tile in Multimodel Chat, and the built-in hosted model is still V4 Flash. Your own key reaches V4.1 Flash through OpenRouter or a custom endpoint, with no markup from us.

Which should you choose

If you…	Best option
Run long agent loops with big, stable prompt prefixes	V4.1 Flash off-peak, cache hits at $0.003
Need the hardest multi-step coding tasks solved	Claude Opus 5, and budget 40× more per session
Want one model for everything at the lowest bill	V4.1 Flash on EU afternoons and weekends
Already have an OpenRouter key and want failover	V4.1 Flash through OpenRouter, pin two providers
Need data processed in a specific region	OpenRouter in-region routing with V4.1 Flash

What you give up

JSON schema enforcement is missing. V4.1 Flash supports response_format for JSON output without schema enforcement. Pipelines that depend on a grammar-constrained decoder should use a model that has one.

Reasoning effort needs per-workload tuning. The 1-to-100 dial defaults to 100, where the model thinks hard and bills for it. Turn it down for classification and extraction, and measure rather than assume the default fits.

Provider mix adds variance. Four providers serve this model on OpenRouter at meaningfully different speeds. One launch-day endpoint ran at 4 tokens per second against the best provider's 142. Pin your providers.

Vendor benchmarks with no neutral score yet. Every number above comes from DeepSeek's framework. They are reproducible from the released repo, more than most labs offer, and they stay DeepSeek's numbers until someone independent publishes theirs.

FAQ

Is DeepSeek V4.1 Flash free? No. The API bills $0.15 per million input tokens on a cache miss, $0.003 on a hit, and $0.60 per million output off-peak, with peak hours doubled. The weights are MIT-licensed, so self-hosting is legal, though a serious deployment needs thousands of GPUs.

What happens to deepseek-v4-pro on September 14, 2026? From 04:00 UTC that day, every V4 Pro request routes to V4.1 Flash at V4.1 Flash rates, until V4.1 Pro ships. No opt-out exists, so teams with pinned models are running comparison evals this week.

Is it cheaper than GPT-5.6 Sol or Claude Opus 5? Per token, yes: roughly 30× on output against Sol's promotional rate and 40× against Opus 5. Cached input widens it, at $0.003 per million against $0.40 and $0.50.

Can I use it in a BYOK chat app? Yes, two ways: add DeepSeek through an OpenRouter connection, or point a custom OpenAI-compatible endpoint at https://api.deepseek.com with deepseek-flash as the model. Both keep provider billing on your own account.

---

Start your free trial → — 7 days, all providers, no credit card required.