DeepSeek V4.1 Flash 1017: the most-called LLM, reviewed with data

Between its launch and early November 2026, one model quietly became the most-called LLM on OpenRouter, the aggregator that routes API traffic for millions of developers: DeepSeek V4.1 Flash 1017. This review looks at what the public usage data actually says, what DeepSeek's own documentation lists, and where the model fits. Every third-party number below carries its source; prices come live from our catalog.

The usage numbers

OpenRouter's weekly leaderboard ranks models by tokens processed, prompt and completion tokens together, with each dated variant counted separately. In the week ending November 5, V4.1 Flash 1017 processed 56.5T tokens, up 33% week over week — more than the next three models combined. In its first full month on the platform it processed 121T tokens. The second bar is the same model's earlier September snapshot, which OpenRouter still ranks on its own.

Counting by author instead of by model, DeepSeek as a whole took 29.8% of text requests on OpenRouter in the last complete week — ahead of OpenAI at 18.8% and Google at 14.4%. Z.ai (the GLM vendor) and Qwen together add another 17.5%.

What the model is, officially

DeepSeek's API documentation is unusually direct about what the name means: the API's deepseek-flash currently maps to DeepSeek-V4.1-Flash-1017, and the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired but still accepted — their requests are served by the 1017 snapshot and billed at the Flash price. That is also why our catalog carries the dated id: a dated snapshot is reproducible, while an alias can change underneath you.

• Context window of 1M tokens; maximum output of 384K.
• Thinking is the default mode; a non-thinking mode is available for lower-latency calls.
• Tool calls, JSON output, an Anthropic-format endpoint and vision input are all supported.
• A concurrency limit of 2500, against 500 for DeepSeek's V4 Pro tier.
• FIM completion (beta) in non-thinking mode, plus chat prefix completion (beta).

Most-used is not the same as smartest

Usage leadership is not a quality claim, and OpenRouter says so itself: token totals "show how much a model is used, not which model is best for a task." For quality, the independent evaluator Artificial Analysis runs an Intelligence Index; the ten leaders currently displayed on OpenRouter's benchmarks tab are all frontier reasoning models, and no Flash-tier model appears among them. Qwen3.8 Max, fifth on that list, is in the price table below.

None of that contradicts the usage data. The Flash tier is not built to top an intelligence index; it is built to keep cost per task low at scale, which is what agent loops, classification and extraction workloads actually optimise for. DeepSeek's own price structure points the same way: two peak windows on UTC weekdays, with nights, weekends and Chinese public holidays all off-peak at half price.

Price and performance

Qwen3.8-Max costs 10.0× as much per output token as DeepSeek V4.1 Flash 1017.

The table reads live prices from our catalog, so it cannot go stale; time-window discounts, where they exist, are listed on each model page. The ratio sentence above is computed from the same live output prices.

Who it fits — and who it does not

• High-frequency agent loops: the 2500-concurrency limit and a cache-read price at a few percent of input reward a stable, repetitive prefix. See how prompt caching cuts the bill.
• Classification, extraction and summarisation pipelines, where the quality bar is already met and volume is the cost driver.
• Chat and support products that need the 1M context window more than frontier reasoning.
• Not a fit: the hardest reasoning or research tasks. The Intelligence Index table above is the better shopping list for those — at a very different price level.

Try it

The model id is the dated string. The request is the same OpenAI-compatible call every other model on this site takes — one key, one endpoint.

curl https://api.modelsite.ai/v1/chat/completions \
  -H "Authorization: Bearer $MODELSITE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash-1017",
    "messages": [{"role": "user", "content": "Explain TCP slow start in two sentences."}]
  }'

Sources

• OpenRouter Rankings — usage data through Nov 5, 2026, published under CC BY 4.0; the charts and tables above are our own rendering of the cited numbers.
• DeepSeek API Docs — Models & Pricing, for the official model mapping, specifications and price structure.
• Artificial Analysis Intelligence Index, top models as displayed on OpenRouter's Benchmarks tab, November 2026.