The Pulse
DeepSeek V4.1 Flash Passes 1 Trillion Tokens on OpenRouter
DeepSeek V4.1 Flash processed more than 1 trillion tokens on OpenRouter shortly after its September 10 launch. The open-weight model combines a 552-billion-parameter mixture-of-experts backbone, native vision and a one-million-token context
AI.info Team ·
“Based on the benchmark results, it does appear DeepSeek V4 is going to be very competitive against its U.S. rivals.”
Lian Jye Su, chief analyst at technology research and advisory firm Omdia, made that assessment in April after DeepSeek introduced its V4 family. Five months later, the company’s newest model is providing a more concrete measure of demand: DeepSeek V4.1 Flash has passed 1 trillion processed tokens on OpenRouter, the model-routing platform, shortly after its September 10 release.
OpenRouter’s public DeepSeek model page shows V4.1 Flash at roughly 1.05 trillion tokens processed. A post attributed to OpenRouter says the model reached 1 trillion tokens within its first 24 hours on the platform and was tracking toward about 2.8 trillion tokens over its first 48 hours. OpenRouter did not say that the figure represented all global usage, and DeepSeek has not published a companywide token count.
The distinction matters. OpenRouter measures traffic routed through its own marketplace, not requests sent directly to DeepSeek or through other providers. Even so, the early volume places V4.1 Flash among the most heavily used new paid models on the service and offers a visible signal of how quickly developers are testing it in production agents, coding tools and long-running workflows.
OpenRouter records a fast start for DeepSeek’s new Flash model
V4.1 Flash arrives on OpenRouter at a moment when model usage is increasingly driven by software agents rather than one-off chat prompts. OpenRouter said in June that agentic requests consume about 15 times more tokens per request than ordinary human interactions. The company also found that DeepSeek’s V4 Flash accounted for 70% of DeepSeek’s agentic token flow by the end of May, roughly one month after the original V4 launch.
The new figures suggest that V4.1 Flash is extending that pattern. OpenRouter’s listing identifies the model as a sparse mixture-of-experts system with a 552-billion-parameter backbone. The model activates 8 billion parameters while processing input and 16 billion while generating output, an arrangement designed to reduce the amount of computation required for long, input-heavy tasks.
OpenRouter’s public page does not provide a full historical time series for the 1.05 trillion-token figure, so the number should not be read as a daily run rate. A separate tracking page published by TKX showed 679 billion tokens routed through OpenRouter during a 24-hour period on September 11, while the OpenRouter listing showed cumulative processed volume above 1 trillion. Those measurements use different windows and should not be treated as contradictory totals.
DeepSeek targets the cost of reading long contexts
DeepSeek’s September 10 announcement focuses less on raw parameter count than on the cost of repeatedly reading large amounts of information. V4.1 Flash uses a Causal Encoder-Decoder architecture with a 20-layer causal encoder followed by a 20-layer decoder. The encoder processes the incoming context, while the decoder generates the response from the encoded representation.
The technical design changes the model’s key-value cache, the memory structure used to retain attention information during inference. DeepSeek says V4.1 Flash reduces the global cache footprint to about 890 bytes per token, roughly one-quarter of the previous V4 Flash figure. The company also says the new model requires one-quarter as much high-bandwidth memory and one-eighth as much solid-state storage for its cache compared with the previous generation.
Those reductions target a specific problem in agent software. An agent may reread the same repository, browser state, documentation set or task history across dozens of steps. If each pass requires storing and retrieving a large context, cache memory and storage traffic can become a larger cost than the model’s headline output price. DeepSeek’s architecture is designed for that repeated-read pattern rather than only for short conversational exchanges.
The model supports contexts of up to 1 million tokens and accepts images natively alongside text. DeepSeek’s model card says it was trained from scratch on a multimodal corpus containing 45 trillion tokens, with sparse attention trained at a 64,000-token sequence length before the context was extended to 1 million tokens at 34 trillion tokens into training.
Pricing helps explain the token surge
DeepSeek launched V4.1 Flash with a peak and off-peak pricing schedule. At off-peak rates, the company lists input at $0.003 per million cached tokens, $0.15 per million uncached input tokens and $0.60 per million output tokens. Peak rates double those prices to $0.006, $0.30 and $1.20 per million tokens, respectively.
OpenRouter’s post says 90% of the first trillion tokens were cache reads, with market pricing of about $0.006 per million tokens. That mix makes the headline volume less surprising: a large agent workload can send the same project files, instructions or tool state through many consecutive requests, creating repeated prefixes that qualify for cache pricing.
DeepSeek has used caching as a cost-control measure since 2024. In an earlier company announcement, the API provider said its service was designed to handle up to 1 trillion tokens per day, with no stated concurrency or rate limits. That figure described the platform’s capacity rather than V4.1 Flash usage, and it predates the new model. The current OpenRouter data is therefore better understood as a usage milestone for one model on one external platform, not proof that DeepSeek has reached 1 trillion tokens per day across its entire business.
OpenRouter’s own pricing and routing system also affects the result. The platform aggregates requests from developers using a unified API and sends them to participating providers. Usage recorded there reflects the platform’s customer base, provider availability, routing choices and model aliases. It gives a useful window into adoption, but it is not an independent census of global inference traffic.
DeepSeek is retiring older V4 endpoints
DeepSeek made V4.1 Flash the default destination for several existing model names when it launched. The company retired V4 Flash and V4 Flash Vision Exp, while requests sent to the older deepseek-v4-flash and deepseek-v4-flash-vision-exp names temporarily route to V4.1 Flash for compatibility.
DeepSeek also plans to route all deepseek-v4-pro requests to V4.1 Flash beginning at 04:00 UTC on September 14, 2026. Those requests will be billed at V4.1 Flash rates until the company releases V4.1 Pro. The change means some traffic attributed to legacy V4 names may soon represent the new model, making precise comparisons between versions harder unless providers preserve model-level records.
Official partners WorkBuddy, including CodeBuddy, and OpenCode support V4.1 Flash. DeepSeek has released the model weights through Hugging Face under the MIT license and published a technical report describing the architecture, cache system and inference setup. The model is large enough to require substantial hardware for full-precision deployment, despite its low active-parameter count during each pass.
The number signals adoption, not model quality
A trillion processed tokens says more about distribution and price than about accuracy. Token volume can rise because developers run longer prompts, agents take more steps, applications retry failed calls or users experiment with a cheap endpoint. None of those measures, by themselves, establish that V4.1 Flash produces better answers than a competing model.
DeepSeek’s own evaluation page lists scores including 90.9 on GPQA Diamond, 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE, though the model card notes that some results depend on the testing harness and tool configuration. DeepSeek says V4.1 Flash outperforms V4 Pro on performance, cost, speed and total runtime, but those comparisons come from the company’s release materials and require independent replication.
The adoption figure still has practical significance. Developers tend to send large volumes through models that combine low prices with acceptable reliability, especially when agents repeat context across many calls. V4.1 Flash’s early OpenRouter traffic suggests that its cache design and pricing are attracting exactly that kind of workload.
For now, the verified claim is narrow: OpenRouter shows more than 1 trillion tokens processed for DeepSeek V4.1 Flash, and OpenRouter attributes a 1 trillion-token first-day milestone to the model. DeepSeek’s broader API capacity, direct traffic and total global usage remain separate figures.