Most-Used AI Models on OpenRouter by Weekly Tokens

DeepSeek V4.1 Flash processed 40.53 trillion tokens on OpenRouter in the week, 39.73 trillion of them prompt tokens, and OpenRouter puts its week-over-week change at +43%.¹ It also led the trailing-thirty-day table with 109T.¹ OpenRouter's model page describes it as a sparse mixture-of-experts model, the first built on DeepSeek's Causal Encoder-Decoder architecture, activating 8B parameters on input and 16B on output from a 552B-parameter backbone, with a 1,048,576-token context window; it was released on Sep 10, 2026.⁶ Its weights are on Hugging Face under the MIT license.¹⁶

Z.ai's GLM 5.3 Flash processed 12.67 trillion tokens, up 26% on the week before by OpenRouter's measure, and ranked third over the trailing thirty days with 55.8T.¹ OpenRouter describes it as a native multimodal model for efficient coding and long-horizon agent tasks, whose hybrid sparse and linear attention keeps long-context behavior accurate while reducing compute; it lists a 1,048,576-token context window and a release date of Aug 26, 2026.⁷ The weights are published on Hugging Face under the MIT license.¹⁶

StepFun's Step 5 Preview reached third place in its first week: OpenRouter lists it as released on Oct 8, 2026, marks it "new" in the weekly table, and counts 12.09 trillion tokens, all but 68 billion of them prompt tokens.¹⁸ On Oct 12 alone it was second, with 3.25T.¹ OpenRouter calls it StepFun's flagship model for agentic work, a sparse mixture-of-experts with 27B active and 600B total parameters and a 1,000,000-token context window, served by a single provider.⁸ No Step 5 weights appear among StepFun's Hugging Face releases, the newest of which is Step-3.7-Flash (May 2026).¹⁷

Xiaomi's MiMo-V2.6-Flash processed 11.54 trillion tokens, up 19% on the week before.¹ OpenRouter's model page calls it an open-source foundation model built on a mixture-of-experts architecture with 309B total parameters and 15B activated per token, using hybrid attention, with native multimodal input and a context window listed at 1,050,000 tokens; it was released on Sep 21, 2026.⁹ The weights OpenRouter links to, MiMo-V2.6-Flash-RL, are on Hugging Face under the MIT license.¹⁶

Tencent's Hy4 preview processed 8.31 trillion tokens, up 31%, and ranked fourth over the trailing thirty days with 39.4T.¹ OpenRouter describes it as a mixture-of-experts model with 49B active parameters out of 770B total, designed for coding agents, complex tool-use workflows and multi-step productivity tasks, with a 1,048,576-token context window; it lists a release date of Aug 28, 2026.¹⁰ Despite the "preview" label, its weights are public on Hugging Face under the Apache-2.0 license.¹⁶ Tencent's earlier Hy3 sits at #13 with 2.78T.¹

This row is only the free variant of NVIDIA's Nemotron 3 Ultra: OpenRouter ranks variants separately, and the paid endpoint isn't in this week's top 20.¹ The free variant processed 7.17 trillion tokens, up 30%.¹ OpenRouter describes the model as a hybrid Transformer-Mamba mixture-of-experts with 55B active parameters out of 550B total, released on Jun 4, 2026, and notes that free-endpoint use is logged by NVIDIA under its trial terms.¹¹ The free endpoint lists a 1,000,000-token context window against 262,144 for the paid one.⁵ The BF16 weights are on Hugging Face, tagged with a custom ("other") license.¹⁶

OpenAI's GPT-6 Luna processed 7.11 trillion tokens, up 14%; among models in the table without weights on Hugging Face, only Step 5 Preview processed more.¹ OpenRouter describes it as the fast, cost-efficient model in the GPT-6 series, positioned below GPT-6 Sol and aimed at high-volume, latency-sensitive work such as chat and classification; it lists a 1,050,000-token context window and a release date of Sep 22, 2026.¹² Its predecessor, GPT-5.6 Luna, is #19 with 1.81T, down 46% on the week.¹

Anthropic's Claude Opus 5.5 processed 5.35 trillion tokens, and OpenRouter puts its week-over-week change at +106%, the largest rise among the top ten models that weren't new that week.¹ OpenRouter describes it as Anthropic's flagship for demanding reasoning, coding and long-horizon agentic work, succeeding Claude Opus 5, with a 1,000,000-token context window and up to 128,000 output tokens; it lists a release date of Sep 22, 2026.¹³ Claude Sonnet 5.5 sits at #14 with 2.65T, up 186%.¹

DeepSeek V4 Flash 0731 processed 4.66 trillion tokens, down 21% on the week before, one of only four models in the top 20 whose usage fell.¹ OpenRouter describes it as a re-post-trained revision and the general-availability release of DeepSeek V4 Flash: a sparse mixture-of-experts with 13B active parameters out of 284B total, released on Jul 31, 2026, with a 1,048,576-token context window.¹⁴ The original April release, listed as DeepSeek V4 Flash 0423, is #12 with 3.10T.¹⁵ Both are MIT-licensed on Hugging Face.¹⁶

Z.ai's GLM 5.3 processed 3.96 trillion tokens, up 33%, putting Z.ai in the top ten twice.¹ OpenRouter describes it as a large-scale reasoning model for complex software engineering and long-horizon agent tasks, with a 1,048,576-token context window; reasoning is always on, with "low", "high" and "max" effort settings and "max" as the default. It lists a release date of Aug 18, 2026.¹⁵ The weights are on Hugging Face with a custom ("other") license tag.¹⁶ The older GLM 5.2 is #20 with 1.53T.¹