DeepSeek waited barely a month after the V4 generation to change the subject. On September 10, 2026 the lab introduced DeepSeek-V4.1-Flash – the smallest model in a brand-new architecture family, with native vision, a 1M-token context window, and open weights under the MIT license. The headline claim is deliberately provocative: the cheap "Flash" model now beats the flagship V4 Pro on agentic benchmarks, while costing roughly a quarter of the price.
I read the launch post, the pricing page, and the technical report's evaluation tables so you don't have to – and added the two things launch posts never give you: the full price math (peak vs off-peak, cache hits, migration dates) and an honest benchmark comparison against today's frontier models – GPT-6 Astra, Claude Fable 5.1, and GLM 5.3.
Source note: specs, prices, and scores below come from DeepSeek's official announcement, the official API pricing page, and the model card's evaluation tables (all runs at maximum reasoning effort). They are vendor-reported numbers, not independent VelsTech benchmarks – and I say exactly where that matters.
The short version
- New architecture, tiny active footprint: 552B total parameters, but only 8B active per token on input (prefill) and 16B on output (decode), thanks to a Causal Encoder-Decoder design plus compressed sparse attention.
- The Flash beats the Pro: V4.1-Flash tops V4-Pro on Terminal-Bench 2.1 (90.6% vs 87.9%), DeepSWE (74.2% vs 62.7%), AutomationBench (54.8% vs 43.2%), and several more agentic tests in DeepSeek's own tables.
- It even takes rounds off closed flagships: best-reported scores on Terminal-Bench 2.1 (90.6%), DeepSWE v1.1 (74.2%), CyberGym (88.1%), HLE with tools (63.9%), AutomationBench (54.8%), and Codeforces (3471) – ahead of every rival in DeepSeek's comparison table.
- Pricing is the real masala: $0.30 / $1.20 per million input / output tokens at peak, halved to $0.15 / $0.60 off-peak, with cache hits at $0.006 ($0.003 off-peak). That is ~33x cheaper input than GPT-6 Astra or Claude Fable 5.1.
-
Live now as
deepseek-flash, with native vision. Olddeepseek-v4-flashnames keep working but route to the new model. V4 Pro was supposed to retire on Sep 14 – DeepSeek has since extended it, so both stay available for now.
What V4.1 Flash actually is
V4.1-Flash is the first and smallest member of DeepSeek's V4.1 family – a multimodal Mixture-of-Experts model that reads images and text and writes text. It was trained from scratch on 45T tokens, supports thinking and non-thinking modes plus a dial (reasoning_effort 1–100) that trades cost for accuracy, and is released open-source (MIT) with weights on Hugging Face.
| Spec | DeepSeek-V4.1-Flash | V4-Flash (prev. gen) | V4-Pro |
|---|---|---|---|
| Total parameters | 552B | 284B | 1.6T |
| Active per token | 8B prefill / 16B decode | 13B | 49B |
| Architecture | Causal Encoder-Decoder, 40 layers (20+20) | MoE + hybrid attention | MoE |
| Experts | 384 routed + 1 shared, 6 active | – | – |
| Context / max output | 1M / 384K | 1M / 384K | 1M / 384K |
| Vision | Native (DeepSeek-ViT) | Separate exp. variant | Not supported |
| KV cache per token | ~890 bytes (FP4) | ~4x larger | – |
| License | MIT (open weights) | MIT | MIT weights / hosted API |
Two engineering ideas do most of the work. Compressed Sparse Attention v2 shares key-value data across layers instead of recomputing it, and SWA Bounded Replay reconstructs short-window attention on the fly instead of storing it – so the persistent cache shrinks to roughly 1/4 of V4-Flash in HBM and 1/8 on SSD, and ~1/437th of the original V1. For agent workloads that re-read huge contexts all day, that is where the money goes.
Price details: the full math
DeepSeek kept peak/off-peak pricing and passed the efficiency savings straight to the rate card. New prices took effect September 10, 2026. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday – everything else is off-peak at exactly half price.
| Per 1M tokens | V4.1 Flash peak | V4.1 Flash off-peak | V4 Pro peak |
|---|---|---|---|
| Input, cache miss | $0.30 | $0.15 | $1.32 |
| Input, cache hit | $0.006 | $0.003 | $0.044 |
| Output | $1.20 | $0.60 | $3.96 |
Three things worth noticing. First, the cache-hit discount is 50x:
repeated system prompts, repo context, and tool definitions cost $0.006/MTok instead
of $0.30. Agents that reuse context – which is nearly all of them – get dramatically
cheaper than the sticker price suggests. Second, V4.1-Flash is
~4.4x cheaper than V4 Pro on input and ~3.3x on output at peak, and
the gap doubles off-peak. Third, the migration is automatic: set the model to
deepseek-flash; legacy deepseek-v4-flash and
deepseek-v4-flash-vision-exp names now route to V4.1-Flash at Flash rates.
One correction to the launch post: it said all deepseek-v4-pro traffic
would move to Flash from 04:00 UTC on September 14. DeepSeek's pricing page now says
V4 Pro stays available with unchanged billing after that date – so
treat the "Pro is phased out" line as superseded. Concurrency limits are 2500 for
Flash vs 500 for Pro, another quiet reason to default to Flash.
Benchmarks vs the frontier: where Flash wins and where it doesn't
The table below is DeepSeek's own comparison (maximum reasoning effort, temperature 1.0 / top_p 0.95, DeepSeek Harness minimal mode for code agents). Rival columns are Opus-5.0, GPT-5.6 Sol, K3, GLM-5.3, V4-Pro, and V4-Flash. Read it as "DeepSeek's best foot forward" – harness and effort settings favor the home team, and independent reruns will follow.
| Benchmark | V4.1 Flash | Best rival in table | Verdict |
|---|---|---|---|
| Terminal-Bench 2.1 (agentic terminal) | 90.6% | 89.1% (Opus-5.0) | Flash leads |
| DeepSWE v1.1 (software engineering) | 74.2% | 74.0% (Opus-5.0) | Flash leads |
| CyberGym | 88.1% | 84.5% (Sol / GLM-5.3) | Flash leads |
| HLE with tools | 63.9% | 63.6% (Opus-5.0) | Flash leads |
| AutomationBench (office workflows) | 54.8% | 50.3% (Opus-5.0) | Flash leads clearly |
| Agent's Last Exam | 31.8% | 28.6% (Opus-5.0) | Flash leads |
| Codeforces rating | 3471 | 3348 (V4-Pro) | Flash leads |
| MathArena Apex | 65.6% (tie K3) | 65.6% (K3) | Tied best |
| GPQA Diamond (reasoning) | 90.9% | 94.1% (GPT-5.6 Sol) | Behind flagships |
| HLE, no tools | 36.8% | 56.3% (Opus-5.0) | Behind – big gap |
| Terminal-Bench 4.0 | 31.2% | 51.8% (Opus-5.0) | Behind newest harness |
| SEC-Bench Pro | 62.8% | 74.3% (GPT-5.6 Sol) | Behind |
| ExploitGym | 15.3% | 33.7% (GPT-5.6 Sol) | Behind |
| ProgramBench / NL2Repo | 20.3% / 64.0% | 37.0% / 75.3% (Opus-5.0) | Behind |
The pattern is consistent: V4.1-Flash dominates the "do long multi-step work" benches – terminal agents, repo-scale coding, automation, tool-using exams – while trailing the biggest closed models on closed-book reasoning (HLE without tools) and the very newest harnesses (Terminal-Bench 4.0, SEC-Bench). Against its own family the story is unambiguous: it beats V4-Pro on nearly every agentic row (TB 2.1, DeepSWE, CyberGym, HLE+tools, AutomationBench, Agent's Last Exam) and on base-model coding (HumanEval 79.4%, BigCodeBench 60.6%).
How does this relate to the newest flagships on this site? GPT-6 Astra (ARC-AGI-3 99.9%, OSWorld 2.0 72.6%, Terminal-Bench 4.0 57.9%) and Claude Fable 5.1 (Terminal-Bench 4.0 55.8%, HLE with tools 65.0%) still lead on their strongest harnesses – but those numbers come from different labs, different harnesses, and different dates, so treat cross-vendor comparisons as directional, not a controlled race. The fair summary: Flash is the first open-weights model that looks genuinely competitive with flagships on agent work rather than just chat.
Frontier price war: Flash vs everyone
Benchmarks are only half the masala. Here is what the same work costs per million tokens (standard/peak rates; verify each vendor's page before committing):
| Model | Input / 1M | Output / 1M | Cache read / 1M |
|---|---|---|---|
| DeepSeek V4.1 Flash (off-peak) | $0.15 | $0.60 | $0.003 |
| DeepSeek V4.1 Flash (peak) | $0.30 | $1.20 | $0.006 |
| DeepSeek V4 Pro (peak) | $1.32 | $3.96 | $0.044 |
| GLM-5.3 | $1.40 | $4.40 | $0.26 |
| Gemini 2.5 Pro (≤200K) | $1.25 | $10.00 | $0.125 |
| GPT-5.6 Sol | $4.00 | $20.00 | $0.40 |
| GPT-6 Astra | $10.00 | $50.00 | $1.00 |
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 |
Put numbers on a realistic agent workload – say ten requests of 200K input tokens at 80% cache hits plus 50K output tokens each, at peak rates (all within the 1M context). On V4.1 Flash that is roughly (0.4M × $0.30) + (1.6M × $0.006) + (0.5M × $1.20) ≈ $0.73 in total. The same workload on a $10/$50 flagship with $1.00 cache reads is roughly $4.00 + $1.60 + $25.00 ≈ $30.60 – about 42x more. Off-peak halves Flash's bill again to about $0.36. Run that through our AI API cost calculator with your own token mix; the break-even vs self-hosting moves a lot at these prices.
What this means if you run local AI
The honest answer first: you are not running 552B at home. The safetensors release is ~763B parameters including its sparsely-accessed memory, and third-party estimates put full-precision serving around 1,160 GB of VRAM at short context – roughly eighteen 80 GB datacenter cards. DeepSeek itself talks about 2,000-GPU-plus-storage deployments for large scale. My 12 GB RX 6800M will not be downloading this one; see how much VRAM LLMs need and the local-LLM GPU guide for what actually fits at home.
What you can use: the weights are MIT-licensed, so expect FP8 originals plus
community GGUF quants, Ollama cloud entries (deepseek-v4.1-flash:cloud
already exists), and vLLM/SGLang support to mature fast. OpenCode and WorkBuddy
(including CodeBuddy) already list V4.1-Flash support – notable for readers of this
site, since OpenCode-style harnesses are exactly where Flash's benchmark strengths
(DeepSWE, Terminal-Bench, AutomationBench) show up. The practical move for local-AI
folks: keep private/sensitive work on your local 7B–35B models, and route the heavy
agentic jobs you used to hesitate over to a $0.15/MTok API instead of a $10 one.
How to try it
- Chat: chat.deepseek.com and the DeepSeek app – fastest way to feel the speed.
- API:
model="deepseek-flash"onhttps://api.deepseek.com(OpenAI- and Anthropic-compatible endpoints). Recommended sampling: temperature 1.0, top_p 0.95, 1M context, max_tokens ≥ 256K for long agents. - Reasoning dial:
reasoning_effort1–100 trades tokens for accuracy – start low for chat, max it for DeepSWE-style tasks. - Open weights: deepseek-ai/DeepSeek-V4.1-Flash on Hugging Face plus the technical report (PDF).
- Vision: no separate vision model anymore – the same Flash endpoint takes images natively.
Caveats before you switch everything
- All instruct scores are max-effort runs. Everyday settings (lower reasoning_effort, different scaffolds) will score lower – and cost less. Both halves of that trade-off are the point.
- Harness matters more than usual here. DeepSeek's code-agent numbers use its own minimal harness at 1M context; Claude Code, Codex, and OpenCode scaffolds score a few points differently on the same model (DeepSWE 65–70 vs 74.2 on mini-SWE).
- Closed-book reasoning still belongs to the giants. HLE without tools (36.8% vs 56.3% for Opus-5.0) and GPQA Diamond (90.9% vs 94.1% for GPT-5.6 Sol) show where the 8B-active diet bites.
- Prices move. Peak windows, the V4-Pro extension, and off-peak halves are policy, not physics – re-check the official pricing page before budgeting.
Bottom line
V4.1-Flash is the rare release that is interesting for two opposite reasons at once. Architecturally, the encoder-decoder + sparse-attention + replay tricks squeeze a 552B model through an 8B-wide straw – and the KV-cache savings (1/4 the HBM, 1/8 the SSD) are genuinely new. Economically, it undercuts every flagship by 30–60x while beating its own Pro sibling where it counts for agent builders. If your workload is long tool-using sessions with big reused contexts, this is the default to beat right now – open weights included. If it is closed-book Olympiad reasoning on the newest harness, keep the flagship subscription warm.