Weekly · AI model usage · week 44, 2026
deepseek-v4.1-flash + 6 more AI models by real usage · week 44, 2026
What developers actually ran this week, by real usage — deepseek-v4.1-flash, glm-5.3-flash, gpt-6-luna, 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 · 58 tokens this week
View on OpenRouter →What we said about deepseek-v4.1-flash
Developers are leaning on this one for high-volume, latency-sensitive work — think batch document processing, log analysis, and agent loops that fire off thousands of calls an hour. That's DeepSeek V4.1 Flash, the sparse mixture-of-experts model that activates 8 billion parameters on input and 16 billion on output, built on DeepSeek's new Causal Encoder-Decoder architecture. And here's what jumps out: 58 trillion tokens this week. That's a staggering number, and it tells you something specific about how teams are using it — not for one-off questions, but for pipelines where that million-token context window lets them stuff entire codebases or document sets into a single pass. The sparse activation keeps the cost-per-token low enough that running it at scale actually pencils out, which is exactly why you see the throughput climbing the way it is. When a model hits volume like this, it usually means developers have wired it deep into automated systems — the kind of workloads that just keep running whether anyone's watching or not.
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 · 15 tokens this week
View on OpenRouter →What we said about glm-5.3-flash
Developers are leaning on this one for long-running agent workflows and high-volume coding pipelines, the kind of tasks where you fire off a job and trust it to hold the thread across hundreds of steps. That model is Z.ai's GLM 5.3 Flash, a native multimodal release built around a hybrid sparse and linear attention design that keeps its footing deep into a one-million-token context. And that architecture choice is exactly why the usage tells a story: fifteen trillion tokens routed through it this week. Numbers at that scale usually mean teams aren't just testing — they're running it in production, feeding it entire codebases, logs, and multi-file refactors without babysitting the window. The "Flash" in the name hints at the draw here: throughput that stays affordable when you're processing enormous volumes, paired with context behavior that doesn't fall apart as the session stretches on. For anyone wiring up autonomous coding agents or batch document work, that combination of reach and efficiency is clearly landing where it counts.
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 · 11 tokens this week
View on OpenRouter →What we said about gpt-6-luna
Developers are leaning on this one for the high-churn stuff — chat backends, classification pipelines, and lightweight agents that fire off thousands of calls a minute. That's GPT-6 Luna, the fast, cost-efficient tier sitting just below Sol in OpenAI's GPT-6 lineup. And the number backing it up is genuinely eye-opening: eleven trillion tokens routed through it this week. That's not a handful of teams experimenting — that's production traffic at serious scale. When you've got a million tokens of context to work with and latency that stays out of your way, it turns out a lot of builders would rather spend Luna on the volume work and save the heavier models for where they actually need them. You can feel the pattern in the usage — people batching classification jobs, powering support flows, running agent loops that don't need the top shelf for every step. Eleven trillion tokens says this slotted neatly into real workloads, and devs kept reaching for it.
Biggest Climber
Step 5 Preview is StepFun's flagship model for agentic work, built on a sparse Mixture-of-Experts architecture (27B active / 600B total parameters). It performs strongly in software engineering and professional
1M ctx · +266 vs last week
View on OpenRouter →What we said about step-5-preview
Developers are wiring this one into agentic pipelines—multi-step coding tasks where the model plans, edits, and iterates across a sprawling codebase without losing the thread. That's Step 5 Preview from StepFun, and the usage curve this week is hard to ignore: up 266% over last week. A big part of that pull is the million-token context window, which lets teams feed in entire repositories, long design docs, and sprawling logs in a single pass instead of chunking everything down. The sparse Mixture-of-Experts setup—27 billion active parameters out of 600 billion total—means people are running serious software-engineering workloads without the cost footprint you'd expect from a model that size. What we're seeing in the numbers is developers leaning on it for the kind of professional, long-horizon tasks that chew through smaller context windows fast. A 266% jump isn't a slow burn—it's folks actively moving real work onto it and apparently liking what comes back enough to keep the requests flowing.
GPT-6.1 Sol is an upgrade to GPT-6 Sol from OpenAI, positioned below the flagship GPT-6 Astra in the GPT-6 series. It is suited for agentic coding, computer use, document-heavy professional
1M ctx · +76 vs last week
View on OpenRouter →What we said about gpt-6.1-sol
Developers are wiring this one into agentic coding loops and computer-use workflows, handing it multi-step tasks and letting it churn through dense professional documents without constant hand-holding. That model is GPT-6.1 Sol, OpenAI's upgrade to GPT-6 Sol, sitting just below the flagship Astra in the GPT-6 lineup. What's catching my eye is the adoption curve: usage jumped seventy-six percent week over week, which is the kind of move you usually see when teams quietly swap it into real pipelines rather than just kicking the tires. The million-token context window is clearly part of the pull — people are feeding it entire codebases and long document sets, the stuff that used to get chopped into awkward chunks. And because it lands below Astra, it tends to be the practical pick for agentic work that runs often and needs to stay affordable. When a jump like this shows up on OpenRouter, it usually means developers found something that fits their day-to-day loop and decided to lean on it.
New On The Board
MiniMax-M3.1-Flash is a multimodal model from MiniMax. Compared with M3, it delivers significant improvements across multiple benchmarks, particularly in complex agentic tasks, coding, and cybersecurity. It is built for complex
1M ctx · 4 tokens this week
View on OpenRouter →What we said about minimax-m3.1-flash
Developers are leaning on this one for sprawling agentic workflows — the kind where an assistant has to chew through a whole codebase, chase down multiple files, and keep its footing across a long chain of tool calls. That's MiniMax M3.1 Flash, and the four trillion tokens it moved this week tells you people aren't just kicking the tires. A big part of the pull is that million-token context window, which means teams can hand it entire repositories or long multimodal sessions without constantly trimming and re-feeding. We're seeing it show up a lot in coding pipelines and security-flavored tasks, where keeping the full picture in view actually matters. Four trillion tokens is a serious volume for a model in the Flash tier — that's the signal that developers are routing real, production-scale traffic through it, not just experiments. When something lands this high on usage, it usually means the latency and context trade-offs are landing right where working teams need them.
Kimi K3.1 Pro is Moonshot AI's flagship open-weight multimodal reasoning model and a post-trained successor to Kimi K3, with a 1M-token context window. It is built for long-horizon coding, 3D
1M ctx · 1 tokens this week
View on OpenRouter →What we said about kimi-k3.1-pro
Developers are feeding this one entire codebases at once — pulling in sprawling monorepos, build logs, and design docs, then asking it to reason across the whole thing without losing the thread. That's Kimi K3.1 Pro from Moonshot AI, and this week it moved a full trillion tokens through OpenRouter. What's driving that number is the kind of work people are pointing it at: long-horizon coding sessions where context actually matters, multi-file refactors, and projects that touch 3D and multimodal inputs alongside plain text. With a million-token window, teams aren't chunking and stitching anymore — they're keeping the whole picture in one place and letting the model stay oriented. A trillion tokens in seven days tells you this isn't experimental poking; these are repeated, production-shaped runs from people who've folded it into their actual workflow. When a model lands that much real traffic, it usually means the long-context promise is holding up under day-to-day pressure, not just in a demo.