Weekly · AI model usage · week 42, 2026
deepseek-v4.1-flash + 6 more AI models by real usage · week 42, 2026
What developers actually ran this week, by real usage — deepseek-v4.1-flash, step-5-preview, 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 · 37 tokens this week
View on OpenRouter →What we said about deepseek-v4.1-flash
Developers are leaning on this one for high-volume pipelines—think log parsing, bulk document summarization, and agentic loops where you're firing off thousands of calls and need them to come back fast and cheap. DeepSeek V4.1 Flash has quietly become a workhorse for exactly that kind of work. The sparse mixture-of-experts design, activating just 8B parameters on input, keeps it light enough to run at scale, and that million-token context window means teams aren't chopping up long inputs just to fit. What really jumps out is the raw adoption: 37 trillion tokens routed through it this week. That's not a handful of teams kicking the tires—that's production traffic, the kind of number you only hit when something's wired into real systems doing real jobs. When developers are pushing that much through a single model, it tells you they've found a reliable slot for it in their stack, somewhere between the quick experiments and the heavy lifting. Flash is earning its place by showing up, call after call.
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 · 19 tokens this week
View on OpenRouter →What we said about step-5-preview
Developers are wiring this one into long-horizon agent loops—the kind where a task kicks off, spawns subtasks, and keeps chewing through a codebase without a human babysitting every step. That's where Step 5 Preview from StepFun is showing up most, and the usage tells the story: nineteen trillion tokens this week. A number like that doesn't come from quick one-off chats. It comes from automated pipelines, agents replaying context over and over, and software engineering workflows that lean on that million-token window to hold an entire repo or a sprawling spec in view. The sparse Mixture-of-Experts design is part of why people keep it in the loop—twenty-seven billion active parameters means it stays responsive even on those heavy, multi-step runs. What stands out here isn't a claim about who's best; it's the sheer volume of real work flowing through it. When teams trust a model to run unattended across big contexts, they run it hard, and this week they clearly did.
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 · 12 tokens this week
View on OpenRouter →What we said about glm-5.3-flash
Developers are wiring this one into long-running coding agents — the kind that churn through a whole repo, plan a refactor, and keep context across dozens of steps without losing the thread. That's where GLM 5.3 Flash from Z.ai is showing up, and the adoption numbers back it up: twelve trillion tokens routed through it this week. That's not a few people kicking the tires. A big part of the pull is the million-token context window, which lets teams feed in sprawling codebases and documentation and still get coherent, grounded responses deep into a session. The hybrid sparse and linear attention setup is clearly resonating with folks running agentic workflows at scale, where you need something that stays efficient when the task horizon stretches out. And because it's natively multimodal, people are leaning on it for jobs that mix code, screenshots, and specs in one pass. Twelve trillion tokens tells you developers aren't just testing this — they're building on it, day in and day out.
Biggest Climber
Claude Haiku 5.5 is Anthropic's small, fast model for high-volume, cost-sensitive work such as summarization, subagents, and browser use. It succeeds Claude Haiku 4.5 with stronger coding, computer use, and
1M ctx · +479 vs last week
View on OpenRouter →What we said about claude-haiku-5.5
Developers are leaning on this one for the high-volume, repetitive stuff that quietly eats their budgets — summarizing long documents, powering subagents, and driving browser-automation loops where every call counts. That's Claude Haiku 5.5, Anthropic's small, fast model built to handle those cost-sensitive jobs without making you think twice about the bill. What jumps out is the usage: up 479% versus last week, one of the sharpest climbs on the board. A big part of that is the million-token context window, which lets people feed in whole codebases or sprawling logs and still get a quick turnaround. Teams are clearly wiring it into pipelines as the workhorse underneath their bigger models — the layer that fans out, fetches, and condenses before anything expensive gets touched. The improved coding and computer-use skills over Haiku 4.5 don't hurt either, especially for folks running agents that actually click around and get things done. When a small, fast model nearly sextuples its traffic in a single week, that tells you developers are moving real production work onto it.
Muse Spark 1.3 is a multimodal reasoning model from Meta for long-running agentic, multi-agent, and coding workflows. It is designed to keep track of information across extended tasks, work through
1M ctx · +431 vs last week
View on OpenRouter →What we said about muse-spark-1.3
Developers are wiring this one into long-running agent pipelines—the kind that spin up sub-agents, hand off tasks, and keep threads straight across hours of work without losing the plot. That's Meta's Muse Spark 1.3, and the usage jump this week is hard to ignore: up 431% over last week. What's driving that? Teams building multi-agent systems need a model that can actually hold context across a sprawling task, and with a one-million-token window, people are feeding it entire codebases, long tool-call histories, and multi-step plans without constantly re-priming it. We're seeing it show up in coding workflows where the model has to remember a decision it made twenty steps back, and in orchestration setups where one agent coordinates several others. A 431% week doesn't happen by accident—it usually means a model crossed some threshold where developers stopped treating it as an experiment and started leaning on it for real pipelines. Worth watching where this lands next week as those agentic workloads mature.
New On The Board
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 · 296 tokens this week
View on OpenRouter →What we said about gemini-4-argon
Developers are leaning on this one for the gnarly stuff — multi-file refactors, agents that chain a dozen steps without losing the thread, and math-heavy work where you actually need the reasoning to hold together. That's Google's Gemini 4 Argon, and it posted 296 billion tokens on OpenRouter this week, which is a serious chunk of real-world traffic. What's driving that volume is the million-token context window: teams are feeding it entire codebases, long design docs, and sprawling conversation histories, then asking it to keep every piece straight across a session. You see it a lot in autonomous agent setups, where a model that drifts or forgets early steps just breaks the whole pipeline. The usage pattern tells the story — this isn't people kicking the tires on a quick one-off. It's long-running jobs, big inputs, and workflows people have wired into their day-to-day. When a number like 296 billion shows up, it usually means something's earned a spot in the toolchain and stayed there.
jev-1.14, on OpenRouter
140 tokens this week
View on OpenRouter →What we said about jev-1.14
Developers are leaning on this one for the long, iterative sessions — the kind where you're refactoring across files, debugging through multiple back-and-forths, and keeping a whole conversation alive without constantly resetting context. That model is jev-1.14, and the usage story here is hard to ignore: 140 billion tokens routed through it this week. That's not a trickle of curious one-off tests; that's teams wiring it into real workflows and letting it run. When a number climbs that high, it usually means the model has earned a spot in people's daily loops — code review, draft generation, pipeline glue, the unglamorous work that actually ships. What stands out to me is the consistency implied by that volume. You don't hit 140 billion on novelty alone; developers keep coming back because it slots cleanly into what they're already building. So whatever jev-1.14 is doing under the hood, the community has clearly decided it's worth feeding real, sustained traffic — and on this chart, that kind of repeat usage is the signal we care about most.