DeepSeek V4.1 Flash launches: faster inference, lower prices, and V4 Pro replacement
DeepSeek has officially released DeepSeek V4.1 Flash, introducing the smallest model in a new architecture family while simultaneously reshaping the economics of its API.
The September 10 launch combines three changes that would normally arrive separately: a new model architecture, native multimodal visual understanding, and a substantial reduction in Flash-series token prices.
DeepSeek is also preparing to redirect requests currently sent to V4 Pro toward V4.1 Flash from September 14, effectively making the cheaper model the temporary production replacement for its higher-tier predecessor until V4.1 Pro arrives.
That makes V4.1 Flash more than another incremental model update: it changes which DeepSeek model developers are likely to use, how much inference costs, and where the company appears to be taking its next architecture generation.
··········
DEEPSEEK V4.1 FLASH STARTS A NEW MODEL ARCHITECTURE.
The new Flash model increases total scale while using a relatively small fraction of its parameters during each processing stage, with native visual understanding built into the architecture.
DeepSeek describes V4.1 Flash as the smallest member of its new architecture family and says the design targets a higher capability ceiling, faster inference, higher throughput, and eventual scaling into larger models.
Technical information accompanying the release indicates a 552-billion-parameter Mixture-of-Experts architecture, significantly larger in total parameter count than the previous V4 Flash generation.
The important number for inference is not simply 552 billion, because the complete model is not activated for every token.
V4.1 Flash reportedly activates roughly 8 billion parameters during prefill and 16 billion during decoding, creating an asymmetric compute profile in which processing the prompt can remain relatively inexpensive while more capacity is allocated when the model generates its response.
This distinction matters for workloads with large prompts, retrieved documents, agent histories, or repeated cached context because total model size does not translate directly into equivalent inference cost per token.
V4.1 Flash also adds native multimodal visual understanding, allowing image information to enter the model architecture rather than requiring a separate vision-specific product path.
........
DeepSeek V4.1 Flash at launch.
Characteristic | DeepSeek V4.1 Flash |
Release | September 10, 2026 |
Architecture | New Mixture-of-Experts architecture |
Total parameters | Approximately 552B |
Active compute during prefill | Approximately 8B parameters |
Active compute during decoding | Approximately 16B parameters |
Input capabilities | Text and native visual input |
Product position | Smallest model in the new V4.1 architecture family |
Primary design targets | Faster inference, higher throughput and larger-model scalability |
........
The architecture therefore appears to pursue a familiar MoE objective more aggressively: maintain a very large pool of specialized parameters while restricting the amount of compute activated for an individual token.
For developers, the practical question will be whether the new routing and multimodal design can preserve these efficiency gains once workloads become long, tool-heavy, or highly agentic rather than simple single-turn prompts.
··········
THE PRICE CUT IS LARGE, AND V4 PRO IS BEING REROUTED.
DeepSeek has reduced Flash pricing at the same time as the architectural upgrade, while V4 Pro is moving toward V4.1 Flash rather than receiving a conventional price reduction.
The new off-peak API price is $0.003 per million cached input tokens, $0.15 per million uncached input tokens, and $0.60 per million output tokens.
Peak-period prices are twice those levels, bringing the corresponding rates to $0.006, $0.30, and $1.20 per million tokens.
Compared with the previous V4 Flash off-peak schedule of $0.007 for cache-hit input, $0.22 for uncached input, and $0.66 for output, the reductions are approximately 57.1%, 31.8%, and 9.1% respectively.
The largest proportional improvement therefore applies to reusable cached context rather than generated output, which can disproportionately benefit applications repeatedly sending large system prompts, documents, repositories, or agent histories.
A workload processing 100 million uncached input tokens and generating 10 million output tokens during off-peak hours would have cost approximately $28.60 under the previous V4 Flash pricing.
At the new V4.1 Flash rates, the same token workload costs approximately $21.00, a reduction of about 26.6% before considering any additional savings from prompt caching.
Using the previous V4 Pro off-peak rates of $0.66 per million uncached input tokens and $1.98 per million output tokens, that same 100M-input/10M-output workload costs approximately $85.80.
Once routed to V4.1 Flash at the new rates, its nominal token bill falls to $21.00, representing a 75.5% reduction for that specific workload configuration.
These Data Studios calculations assume identical token consumption, no cache hits, off-peak execution, and equivalent task completion behavior, so actual savings will depend on how the models differ in reasoning length, retries, tool calls, and output volume.
From September 14 at 12:00 Beijing time, requests targeting V4 Pro are scheduled to be routed to V4.1 Flash and charged at Flash pricing until the future V4.1 Pro becomes available.
This should therefore be understood as a model substitution rather than simply a V4 Pro price cut.
··········
THE BENCHMARK STORY IS STRONG, BUT NOT UNIFORM.
DeepSeek’s published results show particularly large gains in coding, terminal and agentic workloads, although V4 Pro still leads on two of the reported reasoning comparisons.
DeepSeek says V4.1 Flash has surpassed V4 Pro across overall performance, cost, speed, and total completion time, but the individual benchmark results provide a more useful picture than the headline claim alone.
The strongest differences appear in software engineering and terminal-based agent tasks.
Terminal-Bench 3.0 rises from 11.8 for V4 Pro to 30.0 for V4.1 Flash, while DeepSWE v1.1 moves from 62.7 to 74.2 and NL2Repo-Bench increases from 61.5 to 65.4.
Cybersecurity-oriented evaluations also favor the new model in DeepSeek’s reported numbers.
However, V4.1 Flash does not beat V4 Pro on every published comparison.
V4 Pro remains ahead on GPQA Diamond and on the text-only, no-tools subset of Humanity’s Last Exam, which is important when separating conventional difficult reasoning from agentic performance.
........
Selected DeepSeek-reported benchmark results.
Benchmark | V4.1 Flash | V4 Pro 0813 | Higher result |
GPQA Diamond | 90.9 | 92.4 | V4 Pro |
HLE, no tools, text-only subset | 39.1 | 42.7 | V4 Pro |
HLE with tools | 63.9 | 60.0 | V4.1 Flash |
Codeforces rating | 3471 | 3348 | V4.1 Flash |
Terminal-Bench 2.1 | 90.6 | 87.9 | V4.1 Flash |
Terminal-Bench 3.0 | 30.0 | 11.8 | V4.1 Flash |
DeepSWE v1.1 | 74.2 | 62.7 | V4.1 Flash |
NL2Repo-Bench | 65.4 | 61.5 | V4.1 Flash |
CyberGym | 88.1 | 83.3 | V4.1 Flash |
........
The results support a meaningful improvement in the workloads that increasingly matter for coding agents and autonomous software tools, but they should still be treated as vendor-reported launch results rather than a complete independent evaluation.
That distinction is particularly important on release day, when independent testing of latency, reliability, long-running agents, multimodal accuracy, and real production token consumption remains limited.
··········
FOR DEVELOPERS, THE BIGGER CHANGE IS PRODUCT POSITIONING.
Flash is moving from the cheaper member of the DeepSeek lineup toward the default production target during the transition to the full V4.1 family.
A developer selecting between Flash and Pro previously made a relatively conventional trade-off between price, speed, and maximum capability.
The September changes disrupt that hierarchy because V4.1 Flash is not merely being offered alongside V4 Pro: DeepSeek intends to route Pro traffic directly into the new Flash model while the next Pro version is still pending.
Existing V4 Pro users should therefore evaluate V4.1 Flash as a migration even when their application continues calling the same higher-tier endpoint.
Production testing should pay particular attention to response length, latency distributions, tool-call reliability, reasoning behavior, image handling, and agent completion rates because a lower nominal token price does not guarantee a proportionally lower cost per successfully completed task.
Native visual understanding could simultaneously simplify applications that previously needed a separate vision path, particularly document analysis, screenshot interpretation, chart understanding, multimodal agents, and software workflows that operate on graphical interfaces.
The new pricing also changes the economics of repeated-context applications because the deepest percentage cut applies to cache-hit input.
For large-scale agent systems, coding environments, retrieval pipelines, and applications carrying substantial persistent context, that can matter more than a modest reduction in output-token pricing.
DeepSeek has not provided a fixed launch date for V4.1 Pro, meaning the September 14 routing change has a defined start but not yet a defined end.
V4.1 Flash therefore occupies an unusual position: it is nominally the smallest member of a new architecture family, yet it is simultaneously being used to replace the previous generation’s Pro endpoint.
If the architecture scales as DeepSeek intends, the more consequential part of this release may ultimately be what the 552B Flash model reveals about the larger V4.1 models still to come.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
