DeepSeek V4.1-Flash tops OpenRouter's weekly call volume, and TMTPost analysis says this is a victory for cost decisions

By | Shadow Memo

OpenRouter's latest weekly data shows that from October 19 to 25, DeepSeek V4.1-Flash-1017 topped the global rankings with 33.6 trillion tokens in weekly usage, up 801% week-over-week.

Chinese models swept the top five, with combined weekly usage of 93.54 trillion tokens, beating the United States' 22.47 trillion for the 26th consecutive week.

But if you only see the word "first," you will most likely misjudge the true weight of this development.

33.6 trillion tokens is a solid number, but it is not the result of "technological domination"—it is merely the phased realization of a "bills domination."

More precisely, this is a cost decision made by global developers and small and medium-sized enterprises in real workflows, with every API call. Understanding this is far more important than celebrating the ranking itself.

Who is better at doing the math?

V4.1-Flash-1017 is a version released only on October 17. In the same week, usage of the previous generation, V4.1 Flash, fell from a higher level to 18 trillion tokens, down 45% week-over-week.

Users of the old version migrated to the new version, which absorbed almost all of the existing demand, plus a small number of genuinely incremental developers. With one flow out and one flow in, the week-over-week figure was sharply amplified.

The version migration itself shows that users are extremely sensitive to changes in price and performance, that migration costs are low enough, and that switching decisions are almost instantaneous.

That a model can complete such a large-scale old-to-new transition within a week of release precisely proves that its user base is active, engaged, and ready at any time to "vote with their feet."

But it must equally be acknowledged that OpenRouter is not a panoramic view of global large-model usage; it counts traffic routed through this API aggregation gateway.

Vendors' own APIs, enterprise deployments on Azure and Bedrock, private inference, and consumer-facing apps and web products are all outside this pool.

33.6 trillion tokens proves that "on this open aggregation platform, DeepSeek is the model developers are most willing to call," not that "the global AI market is already dominated by DeepSeek."

Only by viewing these two things separately can we continue to ask: why is it, specifically, being called so intensively on this platform?

The answer is not in the top row of the parameter table, but in developers' cost accounting sheets.

V4.1-Flash has 552B total parameters, but in the Prefill stage only 8B are activated per token, and in the Decode stage 16B are activated.

This means its "brain capacity" is close to flagship level, but the computing resources actually used in each inference are pressed to an extremely low level.

Supporting this efficiency is a rather uncommon architectural choice: the 40-layer language backbone is split into a 20-layer causal encoder and a 20-layer Decoder. After the prompt is encoded in the first half, the global KV needed by the second half is projected directly from the end of the encoder, so all historical tokens do not need to repeatedly pass through the entire network.

The official report says Prefill computation is nearly halved.

More critical is the handling of the KV Cache. Agent workflows are characterized by "reading far more than writing": each tool call injects new web pages, code, and terminal output, and the context rapidly expands.

If all history had to be recalculated every time, costs would rise linearly or even exponentially.

DeepSeek's approach is to extract the KV Cache from single-machine memory and place it in an independent 3FS distributed storage cluster, with a cache capacity of about 100 PiB, retaining Agent traffic for 3 days and chat traffic for 7 days.

Repeated context no longer repeatedly burns compute—this is a structural improvement in inference economics in Agent scenarios.

These technical details ultimately point to the same conclusion: between "good enough" and "affordable," DeepSeek has pushed the balance point to a position that a large number of small and medium-sized enterprises cannot refuse.

Some estimates show that the cost for Anthropic's leading model to complete similar Agent tasks is about 28 times higher.

V4.1-Flash-1017 maintains its original price while improving performance, namely $0.15 per million input tokens and $0.60 for output, and is open-sourced under the MIT license.

In an earlier round of pricing adjustments, the cache-hit price for DeepSeek's full API series was cut to one-tenth of the original price, and the cache-hit input price for V4-Flash dropped to 0.02 yuan per million tokens.

This is not "cheapness" in any formal sense, because it is pricing room earned through architectural efficiency.

When a large number of real workflows—such as customer service classification, document tagging, code completion, and Agent tool calls—are far less sensitive to the "absolute upper limit of intelligence" than to the "cost per call," the direction in which global developers vote with tokens is almost predetermined.

The open-source flywheel is harder to replicate than price cuts themselves

If it were merely cheap, DeepSeek's rise to the top could be understood as the short-term result of a price war.

But anyone can fight a price war. What truly turns "cheap" into a structural advantage is that open source has partially shifted pricing power from vendors to users.

On October 20, DeepSeek open-sourced its internal inference framework DeepInfer, compatible with both Nvidia and Huawei Ascend computing platforms.

The significance of this is not that "there is one more open-source project." Before this, even if enterprises obtained the model weights for V3 and R1, that did not mean they could run them with the same efficiency as the official version.

Inference engines, parallelization strategies, cache management, scheduling logic... these are the "tacit knowledge" that turns model weights into low-cost services.

DeepInfer has made this tacit knowledge public: any enterprise can deploy inference services close to official standards on its own cluster, small and medium-sized enterprises no longer need to build optimization capabilities from scratch, and university teams can validate their designs on Ascend.

This forms a cycle that others would find very hard to interrupt with a single price cut: open-source weights lower the barrier to adoption, the open-source inference framework lowers the barrier to deployment, and expanded deployment scale in turn feeds model iteration and cost optimization.

When Chinese models already account for the bulk of global downloads on open-source platforms, and when V4's million-token context and weights are freely open to global developers, the effectiveness of so-called "containment" at the software level continues to decay.

You can ban an institutional account, but it is very hard to ban a set of weights and frameworks that have already entered the toolbox of global developers.

That said, we still need to look at the issue more comprehensively.

After all, more tokens on OpenRouter does not equal more profit.

On the same platform, when sorted by task spending, the top positions in high-value task categories are still occupied by Claude Opus 5.5, GPT-5.6 Sol, and Gemini 3.1 Pro, and the only Chinese model to enter the top ranks of spending on categorized tasks is DeepSeek V4.1 Flash.

For high-complexity, high-risk, high-value workflows, enterprises are still willing to pay a premium for "the smartest one." What DeepSeek has captured is the layer that is largest in volume in real workflows but relatively low in unit value.

This determines that its commercial situation is not as comfortable as its ranking by call volume suggests. Some analyses point out that its actual blended gross margin is far below the normal level of the software industry, and the gap between a team of about 200 people and leading closed-source labs with thousands of employees and tens of billions of dollars in revenue is not something a single climb to the top can close.

API prices have already shown a pullback rhythm at some stages, and "extremely low prices in exchange for scale" is giving way to "how to calculate the unit economics model above scale."

Call volume is part of the moat, but the moat itself does not generate profit.

Questions about technological originality have likewise not disappeared. Allegations of distillation keep recurring, and although DeepSeek's earlier R1 paper responded with the verifiability of its training data sources, the debate over the "degree of originality" of the underlying architecture will not end because of a single weekly No. 1 ranking.

A more realistic constraint comes from computing power: migrating to domestic chips is a step that must be taken, and the cost-effectiveness of a single card can be calculated, but the communication stability at the scale of ten thousand cards and the pretraining efficiency of the next-generation model are the hard variables that determine whether it will still be on the list next year.

Pressure at the geopolitical level is also tightening. The U.S. Department of Homeland Security has required federal civilian agencies to remove DeepSeek products and third-party services developed based on its models from IT systems, expanding the scope of restrictions from the defense and intelligence fields to nearly all federal civilian agencies.

Open source can bypass part of commercial blockade, but it cannot bypass institutional exclusion at the national level.

Conclusion

Over the past year, the share of Chinese models on OpenRouter has been pushed from single digits all the way to a point where they rival overseas models, and the flow of tokens from global developers has undergone a directional change.

The driving force behind this change is not some "crushing" technological breakthrough, but a more plain and also harder-to-replicate-in-the-short-term capability chain: architectural efficiency creates cost space, cost space supports pricing, pricing attracts developers, open source locks developers in as ecosystem participants, and the ecosystem in turn feeds iteration.

But every link in this chain has its boundaries. Leadership in call volume can be won through cost-effectiveness, leadership in profit must be won through irreplaceability, and leadership in technological generations must be won through originality.

DeepSeek won the first exam paper, but that does not mean all three exam papers have been completed.

The 33.6 trillion tokens are real, and the choices of global developers are rational, but beneath the data of "global No. 1," what is truly worth remembering is the reason that made the ranking happen:

In an increasingly expensive industry, someone chose to make "affordability" a core competitiveness.

How far this road can go depends on whether, beyond "cheap," it can continue to offer reasons why "it has to be this one."