On September 10, 2026 DeepSeek released DeepSeek V4.1-Flash — and, unusually, led with the small model. It is a 552B-parameter multimodal mixture-of-experts with a new Causal Encoder-Decoder design that activates just 8B parameters for input and 16B for output, released with MIT-licensed weights, a technical report and a 1M-token context.
DeepSeek says it beats its own V4-Pro on performance, cost, speed and total runtime.
Headline numbers
| Benchmark (DeepSeek-reported, max effort) | V4.1-Flash | V4-Pro-0813 | V4-Flash |
|---|---|---|---|
| DeepSWE v1.1 | 74.2 | 62.7 | 54.4 |
| Terminal-Bench 2.1 | 90.6 | 87.9 | 82.7 |
| Terminal-Bench 4.0 | 31.2 | 12.4 | 7.0 |
| CyberGym | 88.1 | 83.3 | 76.7 |
| AutomationBench | 54.8 | 43.2 | 37.7 |
| HLE (no tools) | 36.8 | 42.7 | 37.8 |
On DeepSWE v1.1 it lands level with Claude Opus 5 (74.0) in DeepSeek’s table, and on Terminal-Bench 2.1 it is ahead of both Opus 5 and GPT-5.6 Sol — though Opus 5 still leads Terminal-Bench 4.0 (51.8 vs 31.2).
Price and routing changes
- API model:
deepseek-flash. V4-Flash and V4-Flash-Vision-Exp are retired; their old names route to V4.1-Flash. - Price: $0.30 input (cache miss) and $1.20 output per million tokens at peak; half that off-peak ($0.15 / $0.60). Peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. DeepSeek says prices came down with this release.
- V4-Pro: DeepSeek first said V4-Pro would route to the new Flash from September 14, then reversed “in response to user demand” and kept it, with billing unchanged. A V4.1-Pro has not been announced yet.
Why it matters
- Efficiency is the story. A KV cache about one quarter the size of the last generation (one eighth on SSD) attacks the cache-hit costs that dominate agent bills.
- Open weights at near-frontier agent scores. An MIT-licensed model matching Opus 5 on DeepSWE v1.1 is a new bar for self-hosted coding agents.
- Native vision. Image input is built in, with DeepSeek reporting Chartography 78.9% and BabyVision 89.6% with tools.
Caveats
- Scores are DeepSeek-reported, with DeepSeek’s own harnesses (and a 72.6 score in its minimal harness against 74.2 in mini-SWE-agent).
- Peak-hour pricing doubles the cost for weekday-morning UTC traffic.
- Closed models still lead the hardest long-horizon tests.
For the wider model landscape, see the 2026 LLM guide.