Why DeepSeek Is Leading OpenRouter: The Economics of Efficient AI

A2Agent Team · 2026-10-28T00:00:00Z

DeepSeek V4.1 Flash holds the top two spots on OpenRouter's weekly usage ranking. The new 1017 build processed 37.3 trillion tokens in the week through October 27, and the original September build added another 17 trillion. Developers send that much traffic to a model when it does the job at a price they can run at scale, and the rest of the top 10 shows the same pattern.

The leaderboard

# Model Provider Weekly tokens Week over week
1 DeepSeek V4.1 Flash 1017 DeepSeek 37.3T +224%
2 DeepSeek V4.1 Flash DeepSeek 17.0T -38%
3 GLM 5.3 Flash Zhipu AI 14.1T +22%
4 GPT-6 Luna OpenAI 9.08T +46%
5 Hunyuan 4 preview Tencent 8.97T +10%
6 MiniMax M3.1 Flash (free) MiniMax 8.71T +75%
7 Claude Opus 5.5 Anthropic 8.07T +19%
8 GLM 5.3 Zhipu AI 5.15T -3%
9 Nemotron 3 Ultra (free) NVIDIA 4.99T +11%
10 MiMo-V2.6-Flash Xiaomi 4.71T -8%

Top 10 models by weekly token volume on OpenRouter, data through October 27, 2026.

The two V4.1 Flash builds together processed 54.3 trillion tokens, about 46% of all the tokens in the top 10. That's more than the next five models combined.

The 1017 build jumped 224% while the September build fell 38%, which looks like users moving from one to the other. Working backward from those percentages, though, the two builds handled roughly 39 trillion tokens the week before and 54 trillion this week. Total DeepSeek V4.1 Flash traffic grew about 40% in a week, on top of the migration.

Seven of the top 10 are Flash-tier, budget or free models. Claude Opus 5.5 and GLM 5.3 are the only flagship-priced models in the top 10, and Opus 5.5 is seventh.

One caveat: OpenRouter ranks models by the prompt and completion tokens they process, so the ranking measures usage, not intelligence. Usage at this scale still says something about what developers find worth paying for, and DeepSeek's release notes explain a lot of it.

What's inside DeepSeek V4.1 Flash

DeepSeek released V4.1 Flash on September 10, 2026, with a stated aim of stronger capability, faster inference and higher throughput on fewer resources. The specs:

DeepSeek V4.1 Flash
Total parameters 552B, mixture of experts
Active parameters About 8B when reading input, 16B when generating output
Architecture New causal encoder-decoder design
Vision Native image understanding in the same model
KV cache 1/4 the HBM and 1/8 the SSD storage of the previous generation
Pricing Peak and off-peak rates; off-peak is 50% of peak
API model ID deepseek-flash

The model has 552B parameters but activates only 8B to read input and 16B to write output, so you get the knowledge of a large model while paying to run a small slice of it on each token. DeepSeek says tests by multiple parties put V4.1 Flash ahead of its own V4 Pro on performance, cost, speed and total runtime, though that's DeepSeek's claim, not an independent result.

The KV cache is the memory a model keeps for context it has already read. V4.1 Flash needs a quarter of the high-bandwidth memory and an eighth of the SSD storage of the last generation for it. DeepSeek notes that cache-hit charges often make up a large share of agent costs, because an agent re-sends the same context, such as a codebase, on every step. A smaller cache makes those repeated reads cheaper to serve.

Vision is built into the main model. A support bot can send a user's error screenshot to deepseek-flash along with the text, with no separate vision model or endpoint.

Off-peak rates are half of peak, so batch jobs like overnight document processing can run at half price just by being scheduled well.

Switching is easy. The old IDs deepseek-v4-flash and deepseek-v4-flash-vision-exp route to V4.1 Flash, and WorkBuddy (including CodeBuddy) and OpenCode support it out of the box.

Why the 1017 build took off

DeepSeek shipped the 1017 build on October 17, ten days before the end of this ranking week. It scores higher than the September build on all 11 benchmarks DeepSeek published, by 2.0 to 11.1 points, at the same token price. DeepSeek also says it finishes tasks faster and writes shorter outputs, so each task costs less than before even though the price per token didn't move. Teams moved their traffic to it within days.

On DeepSWE v1.1, the 1017 build solves 77.0% of tasks for about $0.20 each, ahead of Claude Opus 5 at 74.0% for several dollars a task. Our full breakdown of V4.1 Flash 1017 has the benchmark tables and cost charts.

What an agent task really costs

Agents use models very differently from chatbots. A chatbot answers one question. A coding agent asked to fix a bug reads several files, finds the cause, edits code, runs the tests, reads the errors and tries again, calling the model many times before it's done.

That changes what a model costs. The number that matters is what you pay to finish a task successfully: every call, every retry, all the context re-sent along the way, and the time it takes. A model with a higher score but twice the tokens per task can lose to a cheaper model that finishes in fewer steps. V4.1 Flash is built for that math, with a cheap cache for repeated context, a small active footprint, and, in the 1017 build, shorter outputs.

Try it on A2Agent

Five of the top 10 models on this list are available on A2Agent through one API key: DeepSeek V4.1 Flash 1017, DeepSeek V4.1 Flash, GLM 5.3 Flash, MiniMax M3.1 Flash and GLM 5.3.

DeepSeek V4.1 Flash 1017 costs $0.102 per million input tokens and $0.378 per million output tokens on A2Agent, below DeepSeek's own off-peak rate of $0.15 / $0.60 and well below its peak rate of $0.30 / $1.20, at any hour.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://api.a2agent.me/v1",
)

response = client.chat.completions.create(
    model="deepseek-v4.1-flash-1017",
    messages=[{"role": "user", "content": "Hello"}],
)

print(response.choices[0].message.content)

Because every model sits behind the same key and the same API format, you can run your own tasks through DeepSeek, GLM and MiniMax side by side and compare the cost per finished task directly. New accounts get $5 in free credit to start.

Sources