DeepSeek dropped DeepSeek-V4.1-Flash today, and the version number is doing dishonest work. Reuters flagged the launch this morning; the model card calls it the smallest model in a new architecture family. That is not how point releases talk.
I build WordPress sites and Breakdance / Headspin layouts for a living. My stack already ping-pongs between Claude Fable, GPT Astra, Cursor, and Grok depending on the job. Overseas models only earn a seat if they change cost, speed, or what I can ship before lunch. This one might.
This is not a leak and not a quiet beta. A two-day API test under deepseek-v4.1-flash-expires-on-0910 ended on schedule. The real release is live on DeepSeek’s web app, mobile app, and first-party API, with MIT weights and a technical report on Hugging Face under deepseek-ai/DeepSeek-V4.1-Flash.
The practical box score:
deepseek-flash. Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still accept traffic — but they now resolve to V4.1 Flash at Flash rates.DeepSeek’s Causal Encoder-Decoder split is a 40-layer transformer: 20-layer causal encoder, then 20-layer decoder. The decoder’s global KV cache is projected from the encoder’s final hidden states. Activation lands around 8B parameters per token in prefill and 16B in decode.
That shape is built for the work I actually do — long docs, long tool traces, agent loops — where input dwarfs output. Pair it with FP4 main KV caching (~890 bytes per token, roughly a quarter of V4 Flash), SWA Bounded Replay, and Compressed Sparse Attention 2, and you get a sparsity-and-cache story first, with room to scale the same family later.
Vendor benchmarks (treat them as claims until Artificial Analysis or someone independent measures): GPQA Diamond 90.9, Codeforces 3471, MathArena Apex 65.6, Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2. Same tables still trail GPT-5.6 Sol and Claude Opus 5 on newer Terminal-Bench 3.0 / 4.0 and ProgramBench. Coherent shape: strong on coding, terminal, and agent automation; not the frontier-reasoning king.
Flash pricing itself did not get a cut. Off-peak vs peak still doubles: cache-hit input about $0.003–$0.006 per million tokens, cache-miss $0.15–$0.30, output $0.60–$1.20. Peak windows are 01:00–04:00 and 06:00–10:00 UTC on weekdays.
The cut lands on Pro traffic. After 12:00 Beijing time on September 14, 2026, requests naming deepseek-v4-pro go to V4.1 Flash and bill at Flash rates until V4.1 Pro ships. If your integration hard-codes the Pro model name, nothing “breaks” — you silently get a different model at roughly a third of the prior output price. If you route by capability tier, re-test before that flip.
Claude Fable 5.1 stays my default for long-running Breakdance edits, careful copy, and anything that needs judgment over raw throughput. Cache-read pricing already made Fable the cheap “keep the thread hot” model for multi-hour page builds.
GPT-6 Astra stays the heavy computer-use / hard-reasoning lane when I need the frontier and I’m willing to pay list rates around $10 / $50 per million tokens.
Cursor + Grok stay the IDE loop for fast local refactors and exploratory coding when I want velocity inside the editor.
DeepSeek V4.1 Flash earns a new lane: high-volume, input-heavy agent jobs — bulk HTML-to-section conversions, media alt-text batches, SEO slug passes, WP-CLI script generation — where MIT weights and Flash rates beat paying frontier prices for grunt work. I’m not swapping Fable for it on client-facing design judgment. I am putting it on the eval bench for swarm workers that chew long context and spit structured JSON.
deepseek-v4-pro by name, schedule a capability re-test before September 14 Beijing time.DeepSeek appears to have shipped a generation and declined to call it V5. Annoying for version hygiene. Useful if the architecture holds up under real Breakdance / WordPress workloads. I’ll know more after I burn a few hundred thousand tokens on actual client trees — not another screenshot of a vendor table.