DeepSeek released V4.1 Flash on September 10 -- a sparse mixture-of-experts model with a 552-billion-parameter backbone, activating just 8 billion parameters on input and 16 billion on output. The weights are open, MIT-licensed, and already live on Hugging Face; the context window holds 1,048,576 tokens, and native image understanding is trained into the architecture from the start rather than added afterward.
Pricing is aggressive and shifts by time of day: $0.15 per million input tokens and $0.60 per million output tokens off-peak (01:00-04:00 and 06:00-10:00 UTC on weekdays), doubling to $0.30/$1.20 during peak hours, with a cache-hit input rate as low as $0.003 per million tokens for repeated context. The V4-Flash and V4-Flash-Vision-Exp API identifiers now route to V4.1 Flash automatically, and V4-Pro's own traffic follows on September 14.
DeepSeek's own numbers, run at maximum reasoning effort, show V4.1 Flash ahead of GPT-5.6 Sol and roughly level with Claude Opus 5 on four of five agentic benchmarks it published -- on DeepSWE v1.1, for instance, V4.1 Flash scores 74.2 against GPT-5.6 Sol's 73.0 and Claude Opus 5's 74.0. Those are DeepSeek's own reported figures, not independently reproduced, and they cover a narrow agentic-coding slice, not general capability:
What each benchmark actually covers
- 74.2 vs 73.0 vs 74.0 · DeepSWE v1.1 (agentic coding)
- V4.1 Flash vs. GPT-5.6 Sol vs. Claude Opus 5, DeepSeek's own reported scores
Includes: One narrow coding-agent benchmark, at V4.1 Flash's maximum reasoning effort
Excludes: General reasoning, knowledge, and the broader task mix the Intelligence Index scores - 40 · Artificial Analysis Intelligence Index
- V4.1 Flash's independent score, ranked #6 of 113 models in its class
Includes: A broad suite spanning reasoning, coding, and knowledge tasks, aggregated by an independent evaluator, at the same maximum-effort setting
Excludes: Any of DeepSeek's own agentic-benchmark suite
The architecture behind the price cuts is also new for this generation: a causal encoder-decoder split across 40 transformer layers, which DeepSeek says roughly halves prefill computation on long sequences and cuts persistent KV-cache storage to about an eighth of the previous Flash model's footprint. Those efficiency claims aren't independently benchmarked either, but they're the mechanism DeepSeek points to for why this release can run a much bigger backbone and still charge less.
The gap matters because Artificial Analysis's Intelligence Index -- the only independent measure this newsroom's Scoreboard adopts -- already covers V4.1 Flash at the same maximum-effort setting DeepSeek used for its own claims, and still ranks it behind the two models it's said to beat on a narrower slice of agentic tasks.
- DeepSeek released V4.1 Flash on Sept. 10: 552B parameters, MIT license, 1M-token context.
- Off-peak pricing starts at $0.15 input / $0.60 output per million tokens, doubling at peak.
- DeepSeek's own benchmarks show it beating GPT-5.6 Sol on four of five agentic tasks.
- V4-Pro's API traffic routes to V4.1 Flash starting September 14.
- Caveat: the independent Artificial Analysis index scores it 40 -- below both models it claims to beat.