Weekly · AI model usage · week 38, 2026
gpt-5.6-luna + 6 more AI models by real usage · week 38, 2026
What developers actually ran this week, by real usage — gpt-5.6-luna, deepseek-v4-flash, deepseek-v4.1-flash, and more: the most used, the biggest climbers, and the new entrants, drawn out of the noise.
Most Used
GPT-5.6 Luna is a fast, cost-efficient model in OpenAI's GPT-5.6 series. It is suited for high-volume, latency-sensitive tasks such as chat, classification, and lightweight agentic workflows, providing capable reasoning for
1M ctx · 15 tokens this week
View on OpenRouter →What we said about gpt-5.6-luna
Fifteen trillion tokens this week. That's the number sitting next to GPT-5.6 Luna, and it tells you exactly where this model lives in developers' stacks. Luna is the fast, cost-efficient option in OpenAI's 5.6 series, and the usage pattern lines up with that. People are running it for the high-volume, latency-sensitive stuff — chat interfaces that need to respond instantly, classification pipelines chewing through millions of records, and lightweight agentic workflows where you're firing off a lot of small calls and you can't have cost or lag piling up.
That one-million-token context window is part of the story too. It means teams can feed in big chunks of context without splitting things up, which matters when you're batching at this kind of scale. Fifteen trillion isn't a quality verdict — it's a signal about what developers actually reach for when throughput and budget are on the line.
The full board with every model and its numbers is down in the description. Give it a look, and subscribe so you catch next week's rankings.
DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts model from DeepSeek with 284B total parameters and 13B activated parameters, supporting a 1M-token context window. It is designed for fast inference and
1M ctx · 14 tokens this week
View on OpenRouter →What we said about deepseek-v4-flash
DeepSeek V4 Flash landed on the board hard this week — fourteen trillion tokens flowing through it. That's not a rounding error, that's developers leaning on this thing all day long. And the pattern behind that number makes sense once you look at what it's built for: a Mixture-of-Experts setup, 284 billion total parameters but only 13 billion active per pass. That means fast responses without lighting your budget on fire, which is exactly what teams want when they're running high-volume jobs.
Where's all that traffic coming from? A lot of it looks like bulk work — document processing, log analysis, agent loops that fire constantly. And with a one-million-token context window, folks are dropping entire codebases and long transcripts in without chunking gymnastics. When you're processing at that scale, cheap-and-fast per call adds up quick, and that's the story here.
The full ranked board is down in the description — go see who else made the cut this week. And if you want next week's numbers when they drop, hit subscribe.
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 · 11 tokens this week
View on OpenRouter →What we said about deepseek-v4.1-flash
DeepSeek V4.1 Flash landed hard this week, and the usage numbers tell the story: eleven trillion tokens routed through it. That's not a rounding error — that's developers leaning on it as a daily workhorse.
What are they actually running? A lot of it looks like high-volume pipeline work. That million-token context window means people are feeding it entire repositories, long transcripts, sprawling log files — the kind of jobs where you'd normally chunk everything and stitch it back together. And the "Flash" part matters here: teams are wiring it into batch processing, document extraction, and agent loops where you're firing off thousands of calls and every millisecond compounds.
The sparse mixture-of-experts design — activating a slice of its parameters per request — is clearly the draw for anyone watching their cost-per-token at scale. When you're moving eleven trillion tokens, efficiency stops being a nice-to-have.
The full ranked board is in the description, so go see where this one sits against everything else. And subscribe — we're back next week with fresh numbers.
Biggest Climber
Qwen3.8 Flash is a multimodal reasoning model from Alibaba. It is suited for coding assistance, agentic workflows, visual understanding, document and codebase analysis, desktop interaction, chart analysis, and long-video analysis
1M ctx · +559 vs last week
View on OpenRouter →What we said about qwen3.8-flash
Let's talk about Qwen3.8 Flash, because this one jumped 559 percent week over week — and a spike like that always makes me curious what people are actually plugging it into.
From the usage patterns, developers are leaning on this Alibaba model for a real mix of work. A lot of it is coding assistance and agentic flows — the kind of setups where the model isn't just answering once, but chaining steps, calling tools, and driving a task to completion. The multimodal side is pulling weight too: folks running document and codebase analysis, chart reading, even long-video work.
And that million-token context window is almost certainly part of the story. When you can hand a model an entire codebase or a lengthy document in one go, that unlocks workflows people were previously stitching together by hand.
Whether the jump holds or settles, a 559 percent climb means teams are testing it in production, not just kicking the tires.
Full board's in the description — and hit subscribe so you're here for next week's movers.
Nex-N2.5 is an agentic model built to turn goals into working, verified outcomes. Its core strength is agentic coding within a visual feedback loop: it can explore codebases, implement multi-file
262K ctx · +382 vs last week
View on OpenRouter →What we said about nex-n2.5-pro:free
Let's talk about Nex-N2.5-Pro from Nex AGI, because the usage jump here is hard to ignore — up three hundred and eighty-two percent week over week. That's not a slow climb, that's developers piling in fast.
So what are they actually running it for? This one leans hard into agentic coding — the kind of work where the model explores a codebase, makes multi-file changes, and checks its own outcomes inside a visual feedback loop. That's a workflow a lot of folks have been trying to stitch together by hand, so seeing a model built around it get picked up this quickly tells you people are looking for goals-to-working-code, not just autocomplete.
The two-hundred-and-sixty-two-thousand-token context window is a big part of why it fits that job — enough room to hold a real project in view while it works across files.
And it's running on the free tier right now, which almost certainly fueled that spike. You'll find the full board in the description — subscribe and I'll see you next week with the new numbers.
New On The Board
union-alpha, on OpenRouter
977 tokens this week
View on OpenRouter →What we said about union-alpha
Let's talk about union-alpha, because the usage here genuinely stopped me. Nearly 977 billion tokens flowed through this model on OpenRouter this week — that's not a rounding error, that's developers reaching for it again and again in real workflows.
What are they actually running it for? From the patterns we're seeing, it's landing in the kind of high-volume, repeatable work that eats tokens fast — code generation loops, agent pipelines that call out on every step, batch processing where you're feeding it thousands of requests a day. That scale of adoption usually tells you people have quietly slotted it into production and stopped thinking about it. It just works in the pipe.
And that's the story with these numbers — not what a leaderboard says it should do, but where developers are actually spending their tokens when nobody's watching.
You can see exactly where union-alpha lands on this week's full board — that's all linked in the description below. And if you want next week's rankings the moment they drop, go ahead and subscribe. I'll see you then.
Ling 3.0 Flash VL builds on Ling 3.0 Flash (124B total / 5.5B active MoE from InclusionAI), further strengthening its language capabilities while adding native visual perception and advanced visual
262K ctx · 538 tokens this week
View on OpenRouter →What we said about ling-3.0-flash-vl:free
Let's talk about Ling 3.0 Flash VL from inclusionAI, and the number that caught my eye: 538 billion tokens routed through it this week. That's real developers, real workloads, not a demo spike.
What's pulling that volume? This is a vision-language model, so people are feeding it images alongside text — screenshots, diagrams, documents, UI mockups — and asking it to read, describe, and reason across both. And because it's a mixture-of-experts design with only about 5.5 billion active parameters, it stays light on the compute side, which matters a lot when you're processing that many tokens.
The 262K context window is doing heavy lifting too. Teams are handing it long documents paired with visuals — think manuals, reports, multi-page PDFs — and letting it hold the whole thing at once instead of chunking it up.
Being on the free tier obviously helps adoption, but half a trillion tokens tells you developers are actually building on it, not just kicking the tires.
The full board's in the description. Subscribe and I'll see you next week.