DeepSeek V4.1 Flash 1017: Same Price, Same Weights, Better Model
On 2026-10-17 DeepSeek upgraded DeepSeek V4.1 Flash to DeepSeek-V4.1-Flash-1017 across the web, app and API. The API calling method and the price list did not change: the same deepseek-flash model name reaches the new weights. What changed is the model itself. On Artificial Analysis the Intelligence Index moved from 39 to 45, and the cost per measured task fell from $0.49 to $0.26 — the model got substantially better and cheaper at the same published token price.
Sources: V4.1 Flash,
V4.1 Flash 1017.
The older model page is now marked deprecated; AA still benchmarks its 10k-token workload but no longer updates other results.
What actually changed, and what did not
Not a new model family. Total parameters (552B) and active parameters (16B) are identical, as are the 1M context window and the MIT license. This is a checkpoint upgrade to the same architecture, not a new release with new hardware assumptions.
Not a price change. The published rates are unchanged: $0.15/M input and $0.60/M output off-peak, doubling during the weekday peak windows. DeepSeek states the calling method and pricing remain unchanged for this upgrade.
Better on agentic work, not just on benchmarks. Terminal-Bench 4.0 rose from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0. DeepSeek also reports reduced task runtime and shorter outputs, which is why the cost per task fell even though the token price did not.
The legacy model names now point at it.deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those generations were retired and their requests are served by the 1017 model and billed at the Flash rate.
Vision is now part of the main model. The separate V4 Flash Vision Exp line was retired into V4.1 Flash, which carries native multimodal input. The separate vision record in our dataset reflects a distinct benchmark configuration, not a separate API product.