Weekly · AI model usage · week 43, 2026
deepseek-v4.1-flash + 6 more AI models by real usage · week 43, 2026
What developers actually ran this week, by real usage — deepseek-v4.1-flash, minimax-m3.1-flash:free, glm-5.3-flash, and more: the most used, the biggest climbers, and the new entrants, drawn out of the noise.
Most Used
DeepSeek V4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on
1M ctx · 45 tokens this week
View on OpenRouter →What we said about deepseek-v4.1-flash
Developers are leaning on this one for high-volume pipeline work — document processing, log summarization, batch extraction, and agent loops where the token meter never stops running. DeepSeek V4.1 Flash is pulling that traffic, and the number behind it is genuinely striking: forty-five trillion tokens this week. That's the kind of throughput you only see when a model lands in production jobs, not weekend experiments. The pull here makes sense when you look at the shape of it — a sparse mixture-of-experts design that only activates a slice of its parameters per pass, paired with a million-token context window that lets teams throw entire codebases or long transcripts at it in one go. For anyone running cost-sensitive, large-scale workloads, that combination is exactly what gets wired into a nightly cron job and left to run. Forty-five trillion tokens says developers aren't just testing this — they've handed it the heavy, repetitive work and walked away trusting it to keep going.
minimax-m3.1-flash:free, on OpenRouter
13 tokens this week
View on OpenRouter →What we said about minimax-m3.1-flash:free
Developers are leaning on this one for high-throughput background work—batch classification, log summarization, and the kind of always-on pipelines where cost per token matters more than anything. That model is minimax-m3.1-flash:free, and the usage story here is honestly hard to ignore. Thirteen trillion tokens in a single week puts it in rarefied territory, and the pattern behind that number tells you a lot. When something's free and fast, people stop rationing it—they wire it into cron jobs, let it chew through queues overnight, and route the easy requests to it before escalating to heavier models. That's exactly the shape of this traffic: enormous volume spread across a lot of small, repetitive tasks. You don't hit thirteen trillion tokens with a handful of power users; you get there when it quietly becomes the default workhorse for teams who need throughput at scale. It's the unglamorous backbone work, and clearly a lot of developers have decided this is the tool they're reaching for.
GLM-5.3-Flash is a native multimodal model from Z.ai. It is suited for efficient coding and long-horizon agent tasks. Its hybrid sparse and linear attention architecture maintains accurate long-context behavior while
1M ctx · 11 tokens this week
View on OpenRouter →What we said about glm-5.3-flash
Developers are wiring this one into agent pipelines that have to stay coherent across long, multi-step runs—think code refactors that span dozens of files and workflows that keep context alive over hours. That's where GLM 5.3 Flash from Z.ai is showing up, and the adoption tells the story: eleven trillion tokens routed through it this week. A lot of that volume comes down to its roomy one-million-token context window, which lets teams feed in entire repositories or long task histories without constantly trimming and re-summarizing. The native multimodal side means people aren't just passing text—they're folding in diagrams, screenshots, and mixed inputs as part of the same agent loop. And the "Flash" label matters here: this is being picked for high-throughput jobs where you're firing off a huge number of calls and efficiency per token actually moves your bill. When a model crosses eleven trillion tokens in a single week, that's not a few people kicking the tires—that's production traffic, developers leaning on it for the heavy, repetitive, long-horizon work.
Biggest Climber
jev-1.14, on OpenRouter
+834 vs last week
View on OpenRouter →What we said about jev-1.14
Developers are leaning on this one for high-volume, repetitive pipeline work — the kind of tasks you fire off hundreds of times a day and need to behave predictably every single run. That model is jev-1.14, and the reason it lands on the board this week is pure momentum: usage jumped eight hundred and thirty-four percent over the previous week. That's not a slow climb, that's a sharp turn, and it usually means a handful of teams quietly wired it into something that scales. When a model's traffic multiplies like that in seven days, it's rarely about hype threads — it's about someone finding it reliable enough to put in a loop and walk away. We're seeing it show up in automation chains and batch jobs, places where consistency matters more than flash. Worth keeping an eye on whether this spike holds or settles once the early adopters finish their testing. For now, jev-1.14 earned its spot the honest way — real requests, real developers, a real jump in the numbers.
Gemini 4 Argon is Google's flagship reasoning model, built for complex software engineering, deep mathematical and scientific problem solving, and autonomous multi-step agents. Google trained it to keep context across
1M ctx · +238 vs last week
View on OpenRouter →What we said about gemini-4-argon
Developers are pointing this one at the gnarly stuff — multi-file refactors, long-running agent loops, and the kind of mathematical and scientific reasoning that usually falls apart halfway through. Gemini 4 Argon is Google's flagship reasoning model, and what's catching my eye this week is the trajectory: usage jumped two hundred and thirty-eight percent over last week. That's not a slow adoption curve, that's teams reaching for it deliberately. A lot of that comes down to the million-token context window, which lets people keep an entire codebase or a sprawling research thread in frame without constantly re-feeding chunks back in. You see it showing up in autonomous agents that need to hold a plan across dozens of steps, and in engineering workflows where losing the thread means losing the whole task. A jump that steep usually means people tried it on something real, it held together, and they came back. On a usage board, that kind of week-over-week move is exactly what tells you something's shifting underneath.
New On The Board
NVIDIA Nemotron 3.5 Super VL is an open multimodal mixture-of-experts model from NVIDIA, with 12B active parameters out of 120B total. It takes text, images, documents, GUI screenshots, and video
1M ctx · 1 tokens this week
View on OpenRouter →What we said about nemotron-3.5-super-vl:free
Developers are feeding this one everything — screenshots, scanned documents, product photos, even GUI captures from the apps they're automating. That's NVIDIA's Nemotron 3.5 Super VL, a multimodal mixture-of-experts model running 12 billion active parameters out of 120 billion total, and the usage this week tells a clear story: a full trillion tokens. For a vision-language model, that's a striking amount of traffic, and it points to real pipelines, not just curiosity. A lot of that volume comes from people wiring it into document extraction, interface-reading agents, and workflows that chew through image-heavy data at scale. The million-token context window is clearly part of the draw here — it lets developers push long documents and multi-page captures through in a single pass instead of stitching chunks together. And since it's offered free on OpenRouter right now, the barrier to just trying it on a real multimodal task drops to basically nothing. When a model this specialized crosses a trillion tokens in a week, that's developers voting with their workloads.
GPT-6 Luna is the fast, cost-efficient model in OpenAI's GPT-6 series, positioned below GPT-6 Sol. It is suited for high-volume and latency-sensitive workloads such as chat, classification, and lightweight agentic
1M ctx · 90 tokens this week
View on OpenRouter →What we said about gpt-6-luna:batch
Developers are pushing this one through enormous queues of classification jobs, chat endpoints, and the kind of lightweight agent loops that fire thousands of times an hour. GPT-6 Luna, running here in its batch configuration, is the fast, cost-efficient tier sitting just below GPT-6 Sol in OpenAI's lineup — and the usage tells the story. Ninety billion tokens this week. That's not a model people are babysitting through one careful prompt at a time; that's a model teams are wiring into production pipelines where latency and cost-per-call actually move the budget. The million-token context window gives them room to batch big documents or long transcripts without chunking gymnastics, which explains why so much high-volume work is landing here instead of on a heavier sibling. When you're processing at this scale, the boring stuff — predictable throughput, cheap tokens, enough headroom to not think about it — is exactly what wins the integration. And ninety billion tokens says a lot of developers have already made that call.